{"id":902,"date":"2025-09-14T06:30:14","date_gmt":"2025-09-14T10:30:14","guid":{"rendered":"https:\/\/infohatch.net\/ai\/?p=902"},"modified":"2025-09-14T06:38:50","modified_gmt":"2025-09-14T10:38:50","slug":"building-truly-agentic-ai-systems-2","status":"publish","type":"post","link":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/","title":{"rendered":"Building Truly Agentic AI Systems"},"content":{"rendered":"<p data-start=\"38\" data-end=\"263\"><span style=\"font-size: 14pt;\"><em data-start=\"38\" data-end=\"263\">Peek behind the curtain to reveal the hidden dynamics of developing agentic AI systems\u2014plus hands-on \u201chacks\u201d to boost autonomy and efficiency with the latest advances in reinforcement learning and multi-agent collaboration.<\/em><\/span><\/p>\n<h2 data-start=\"270\" data-end=\"290\"><span style=\"font-size: 18pt; color: #3366ff;\"><strong>Executive Summary<\/strong><\/span><\/h2>\n<p data-start=\"292\" data-end=\"521\"><span style=\"font-size: 14pt;\">Agentic AI \u2260 \u201ca chatbot with tools.\u201d Truly agentic systems can <strong data-start=\"355\" data-end=\"365\">decide<\/strong>, <strong data-start=\"367\" data-end=\"374\">act<\/strong>, <strong data-start=\"376\" data-end=\"385\">learn<\/strong>, and <strong data-start=\"391\" data-end=\"405\">coordinate<\/strong>\u2014under uncertainty, across long horizons, with measurable outcomes and guardrails. The current best practice blends:<\/span><\/p>\n<ul data-start=\"523\" data-end=\"1297\">\n<li data-start=\"523\" data-end=\"638\">\n<p data-start=\"525\" data-end=\"638\"><span style=\"font-size: 14pt;\"><strong data-start=\"525\" data-end=\"547\">Reasoning + acting<\/strong> loops (plan, act, observe, reflect) with tool-use. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2210.03629?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"639\" data-end=\"745\">\n<p data-start=\"641\" data-end=\"745\"><span style=\"font-size: 14pt;\"><strong data-start=\"641\" data-end=\"665\">Search over thoughts<\/strong> (trees\/graphs), not just single chains. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2305.10601?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><span class=\"flex h-4 w-full items-center justify-between absolute\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">NeurIPS Proceedings<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"746\" data-end=\"876\">\n<p data-start=\"748\" data-end=\"876\"><span style=\"font-size: 14pt;\"><strong data-start=\"748\" data-end=\"772\">Purpose-built memory<\/strong> and <strong data-start=\"777\" data-end=\"798\">stateful runtimes<\/strong> (episodic\/semantic\/procedural state). <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2310.08560?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"877\" data-end=\"1043\">\n<p data-start=\"879\" data-end=\"1043\"><span style=\"font-size: 14pt;\"><strong data-start=\"879\" data-end=\"905\">Reinforcement learning<\/strong> that increasingly reduces reliance on supervised labels (RL from AI feedback; pure-RL reasoning). <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2309.00267?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><span class=\"flex h-4 w-full items-center justify-between absolute\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">OpenReview<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"1044\" data-end=\"1169\">\n<p data-start=\"1046\" data-end=\"1169\"><span style=\"font-size: 14pt;\"><strong data-start=\"1046\" data-end=\"1075\">Multi-agent collaboration<\/strong> frameworks and debate\/judge patterns for reliability. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2308.08155?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><span class=\"flex h-4 w-full items-center justify-between absolute\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">Microsoft GitHub<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"1170\" data-end=\"1297\">\n<p data-start=\"1172\" data-end=\"1297\"><span style=\"font-size: 14pt;\"><strong data-start=\"1172\" data-end=\"1200\">Task-grounded evaluation<\/strong> (e.g., SWE-bench) and sandboxed Agent-Computer Interfaces. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/www.swebench.com\/?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">SWE-Bench<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<p data-start=\"1299\" data-end=\"1428\"><span style=\"font-size: 14pt;\">The rest of this report turns those ideas into an implementable blueprint\u2014complete with low-friction \u201chacks\u201d you can apply today.<\/span><\/p>\n<h2 data-start=\"1435\" data-end=\"1475\"><span style=\"font-size: 18pt; color: #3366ff;\"><strong>1) Hidden dynamics of agentic systems<\/strong><\/span><\/h2>\n<h3 data-start=\"1477\" data-end=\"1511\"><span style=\"font-size: 14pt; color: #800000;\"><strong>1.1 Reasoning\u2013acting co-design<\/strong><\/span><\/h3>\n<p data-start=\"1512\" data-end=\"1949\"><span style=\"font-size: 14pt;\">Agents that interleave <strong data-start=\"1535\" data-end=\"1558\">thinking with doing<\/strong> outperform those that separate the two. <\/span><\/p>\n<p data-start=\"1512\" data-end=\"1949\"><span style=\"font-size: 14pt;\">The ReAct paradigm shows agents should <strong data-start=\"1638\" data-end=\"1679\">explain their reasoning, take actions<\/strong>, read observations, and update plans inside one loop\u2014reducing hallucinations and enabling tool-driven information gathering. <\/span><\/p>\n<p data-start=\"1512\" data-end=\"1949\"><span style=\"font-size: 14pt;\">Design implication: your runtime must log <em data-start=\"1847\" data-end=\"1853\">both<\/em> internal thoughts and external acts as first-class state. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2210.03629?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<h3 data-start=\"1951\" data-end=\"1992\"><span style=\"color: #800000;\"><strong><span style=\"font-size: 14pt;\">1.2 Structured search over thoughts<\/span><\/strong><\/span><\/h3>\n<p data-start=\"1993\" data-end=\"2378\"><span style=\"font-size: 14pt;\">Single-path chain-of-thought often gets stuck. <\/span><\/p>\n<p data-start=\"1993\" data-end=\"2378\"><span style=\"font-size: 14pt;\">Moving to <strong data-start=\"2050\" data-end=\"2070\">Tree-of-Thoughts<\/strong> (branching + self-evaluation) and then <strong data-start=\"2110\" data-end=\"2131\">Graph-of-Thoughts<\/strong> (arbitrary DAGs with merges and feedback) unlocks better global choices and cost control (prune weak branches early). <\/span><\/p>\n<p data-start=\"1993\" data-end=\"2378\"><span style=\"font-size: 14pt;\">Build \u201cthought operators\u201d (e.g., <em data-start=\"2283\" data-end=\"2319\">expand, evaluate, backtrack, merge<\/em>) as reusable skills. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2305.10601?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><span class=\"flex h-4 w-full items-center justify-between absolute\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">NeurIPS Proceedings<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<h3 data-start=\"2380\" data-end=\"2433\"><span style=\"color: #800000;\"><strong><span style=\"font-size: 14pt;\">1.3 Memory is an OS problem, not a prompt trick<\/span><\/strong><\/span><\/h3>\n<p data-start=\"2434\" data-end=\"2794\"><span style=\"font-size: 14pt;\">Long-horizon autonomy needs <strong data-start=\"2462\" data-end=\"2479\">tiered memory<\/strong>: transient scratchpads, episodic logs, semantic summaries, and skill libraries\u2014with an <strong data-start=\"2567\" data-end=\"2596\">LLM-driven memory manager<\/strong> deciding what to page in\/out. <\/span><\/p>\n<p data-start=\"2434\" data-end=\"2794\"><span style=\"font-size: 14pt;\">MemGPT formalizes this with interrupts and virtual memory. Your platform should expose memory as a contract, not an afterthought. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2310.08560?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<h3 data-start=\"2796\" data-end=\"2835\"><span style=\"color: #800000;\"><strong><span style=\"font-size: 14pt;\">1.4 Agents are graphs with cycles<\/span><\/strong><\/span><\/h3>\n<p data-start=\"2836\" data-end=\"3106\"><span style=\"font-size: 14pt;\">Production agents revisit goals, loop for tool calls, and branch across roles. <\/span><\/p>\n<p data-start=\"2836\" data-end=\"3106\"><span style=\"font-size: 14pt;\">Libraries like <strong data-start=\"2930\" data-end=\"2943\">LangGraph<\/strong> make the \u201cagent = state machine\/graph\u201d idea explicit, adding human-in-the-loop checkpoints and moderation at specific edges. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/www.langchain.com\/langgraph?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">LangChain<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<h2 data-start=\"3113\" data-end=\"3150\"><span style=\"font-size: 18pt; color: #3366ff;\"><strong>2) Architecture patterns that work<\/strong><\/span><\/h2>\n<h3 data-start=\"3152\" data-end=\"3190\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"3152\" data-end=\"3188\">Core loop (minimal viable agent)<\/strong><\/span><\/h3>\n<ol data-start=\"3191\" data-end=\"3422\">\n<li data-start=\"3191\" data-end=\"3235\">\n<p data-start=\"3194\" data-end=\"3235\"><span style=\"font-size: 14pt;\"><em data-start=\"3194\" data-end=\"3204\">Perceive<\/em> (ingest instruction + state)<\/span><\/p>\n<\/li>\n<li data-start=\"3236\" data-end=\"3284\">\n<p data-start=\"3239\" data-end=\"3284\"><span style=\"font-size: 14pt;\"><em data-start=\"3239\" data-end=\"3245\">Plan<\/em> (draft steps \/ search over thoughts)<\/span><\/p>\n<\/li>\n<li data-start=\"3285\" data-end=\"3310\">\n<p data-start=\"3288\" data-end=\"3310\"><span style=\"font-size: 14pt;\"><em data-start=\"3288\" data-end=\"3293\">Act<\/em> (tool\/API\/env)<\/span><\/p>\n<\/li>\n<li data-start=\"3311\" data-end=\"3343\">\n<p data-start=\"3314\" data-end=\"3343\"><span style=\"font-size: 14pt;\"><em data-start=\"3314\" data-end=\"3323\">Observe<\/em> (results, errors)<\/span><\/p>\n<\/li>\n<li data-start=\"3344\" data-end=\"3390\">\n<p data-start=\"3347\" data-end=\"3390\"><span style=\"font-size: 14pt;\"><em data-start=\"3347\" data-end=\"3356\">Reflect<\/em> (critic\/judge\/Reflexion memory)<\/span><\/p>\n<\/li>\n<li data-start=\"3391\" data-end=\"3422\">\n<p data-start=\"3394\" data-end=\"3422\"><span style=\"font-size: 14pt;\"><em data-start=\"3394\" data-end=\"3422\">Decide to stop or continue<\/em><\/span><\/p>\n<\/li>\n<\/ol>\n<h3 data-start=\"3424\" data-end=\"3453\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"3424\" data-end=\"3451\">Best-in-class scaffolds<\/strong><\/span><\/h3>\n<ul data-start=\"3454\" data-end=\"3999\">\n<li data-start=\"3454\" data-end=\"3540\">\n<p data-start=\"3456\" data-end=\"3540\"><span style=\"font-size: 14pt;\"><strong data-start=\"3456\" data-end=\"3465\">ReAct<\/strong> for tight think\u2013act integration. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2210.03629?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"3541\" data-end=\"3688\">\n<p data-start=\"3543\" data-end=\"3688\"><span style=\"font-size: 14pt;\"><strong data-start=\"3543\" data-end=\"3551\">LATS<\/strong> (Language Agent Tree Search) to unify planning\/acting\/reasoning with a tree search controller. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2310.04406?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"3689\" data-end=\"3856\">\n<p data-start=\"3691\" data-end=\"3856\"><span style=\"font-size: 14pt;\"><strong data-start=\"3691\" data-end=\"3704\">Reflexion<\/strong> to let agents <strong data-start=\"3719\" data-end=\"3736\">self-critique<\/strong> and store distilled lessons for the next episode\u2014weight updates not required. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2303.11366?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"3857\" data-end=\"3999\">\n<p data-start=\"3859\" data-end=\"3999\"><span style=\"font-size: 14pt;\"><strong data-start=\"3859\" data-end=\"3870\">AutoGen<\/strong> to compose multiple specialized agents that converse, use tools, and escalate to humans. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2308.08155?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<h2 data-start=\"4006\" data-end=\"4079\"><span style=\"font-size: 18pt; color: #3366ff;\"><strong>3) Little-known \u201chacks\u201d that immediately improve autonomy &amp; efficiency<\/strong><\/span><\/h2>\n<ol data-start=\"4081\" data-end=\"6451\">\n<li data-start=\"4081\" data-end=\"4327\">\n<h3 data-start=\"4084\" data-end=\"4121\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"4084\" data-end=\"4119\">Action chunking with guardrails<\/strong><\/span><\/h3>\n<ul data-start=\"4125\" data-end=\"4327\">\n<li data-start=\"4125\" data-end=\"4327\">\n<p data-start=\"4127\" data-end=\"4327\"><span style=\"font-size: 14pt;\">Convert multi-step plans into <strong data-start=\"4157\" data-end=\"4175\">chunked macros<\/strong> (\u201cskills\u201d). Maintain a skill registry with success stats; prefer proven macros first. Pair each macro with a <strong data-start=\"4285\" data-end=\"4298\">validator<\/strong> (regex, schema, unit tests).<\/span><\/p>\n<\/li>\n<\/ul>\n<\/li>\n<li data-start=\"4329\" data-end=\"4610\">\n<h3 data-start=\"4332\" data-end=\"4370\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"4332\" data-end=\"4368\">Self-competition for reliability<\/strong><\/span><\/h3>\n<ul data-start=\"4374\" data-end=\"4610\">\n<li data-start=\"4374\" data-end=\"4610\">\n<p data-start=\"4376\" data-end=\"4610\"><span style=\"font-size: 14pt;\">Run <em data-start=\"4380\" data-end=\"4383\">k<\/em> lightweight candidates (self-consistency) then use an internal <strong data-start=\"4447\" data-end=\"4456\">judge<\/strong> to pick the best\u2014works wonders on math, code, and retrieval tasks while bounding cost by adaptive early-stopping. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2203.11171?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<\/li>\n<li data-start=\"4612\" data-end=\"4891\">\n<h3 data-start=\"4615\" data-end=\"4656\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"4615\" data-end=\"4654\">Reflexion memories as few-shot fuel<\/strong><\/span><\/h3>\n<ul data-start=\"4660\" data-end=\"4891\">\n<li data-start=\"4660\" data-end=\"4891\">\n<p data-start=\"4662\" data-end=\"4891\"><span style=\"font-size: 14pt;\">Persist \u201cwhat I learned\u201d lines (failures, fixes, invariants). Feed them back as <strong data-start=\"4742\" data-end=\"4762\">episodic anchors<\/strong> at the next attempt. (Think: \u201cWhen installing Python deps in repo X, pin uvicorn&lt;0.30.\u201d) <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2303.11366?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<\/li>\n<li data-start=\"4893\" data-end=\"5157\">\n<h3 data-start=\"4896\" data-end=\"4942\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"4896\" data-end=\"4940\">Debate + judge only when entropy is high<\/strong><\/span><\/h3>\n<ul data-start=\"4946\" data-end=\"5157\">\n<li data-start=\"4946\" data-end=\"5157\">\n<p data-start=\"4948\" data-end=\"5157\"><span style=\"font-size: 14pt;\">Multi-agent debate boosts accuracy, but it isn\u2019t free. Use a <em data-start=\"5009\" data-end=\"5032\">disagreement detector<\/em> (e.g., variance across candidates) to <strong data-start=\"5071\" data-end=\"5088\">conditionally<\/strong> trigger debate with a judge. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2305.14325?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<\/li>\n<li data-start=\"5159\" data-end=\"5373\">\n<h3 data-start=\"5162\" data-end=\"5198\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"5162\" data-end=\"5196\">Graph-of-Thoughts cost shaping<\/strong><\/span><\/h3>\n<ul data-start=\"5202\" data-end=\"5373\">\n<li data-start=\"5202\" data-end=\"5373\">\n<p data-start=\"5204\" data-end=\"5373\"><span style=\"font-size: 14pt;\">Assign <strong data-start=\"5211\" data-end=\"5235\">budget per node type<\/strong> (expand vs. verify vs. merge); prune weak branches early; reuse evaluated subgraphs across tasks. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2308.09687?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<\/li>\n<li data-start=\"5375\" data-end=\"5635\">\n<h3 data-start=\"5378\" data-end=\"5427\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"5378\" data-end=\"5425\">Lookahead decoding to speed up long outputs<\/strong><\/span><\/h3>\n<ul data-start=\"5431\" data-end=\"5635\">\n<li data-start=\"5431\" data-end=\"5635\">\n<p data-start=\"5433\" data-end=\"5635\"><span style=\"font-size: 14pt;\">For verbose agents (docs\/code), <strong data-start=\"5465\" data-end=\"5478\">lookahead<\/strong> decoding provides <strong data-start=\"5497\" data-end=\"5506\">exact<\/strong> speed-ups by parallelizing n-gram verification\u2014cutting latency without hurting accuracy. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/lmsys.org\/blog\/2023-11-21-lookahead-decoding\/?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">LMSYS<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<\/li>\n<li data-start=\"5637\" data-end=\"5921\">\n<h3 data-start=\"5640\" data-end=\"5689\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"5640\" data-end=\"5687\">Agent-Computer Interface (ACI) &gt; shell-only<\/strong><\/span><\/h3>\n<ul data-start=\"5693\" data-end=\"5921\">\n<li data-start=\"5693\" data-end=\"5921\">\n<p data-start=\"5695\" data-end=\"5921\"><span style=\"font-size: 14pt;\">Provide structured OS-level affordances (open\/edit file, run tests, browse docs, start server) rather than raw bash. Research shows ACI materially improves code-fix success on SWE-bench. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/papers.nips.cc\/paper_files\/paper\/2024\/file\/5a7c947568c1b1328ccc5230172e1e7c-Paper-Conference.pdf?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">NeurIPS Papers<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<\/li>\n<li data-start=\"5923\" data-end=\"6188\">\n<h3 data-start=\"5926\" data-end=\"5966\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"5926\" data-end=\"5964\">Auto-curricula and skill libraries<\/strong><\/span><\/h3>\n<ul data-start=\"5970\" data-end=\"6188\">\n<li data-start=\"5970\" data-end=\"6188\">\n<p data-start=\"5972\" data-end=\"6188\"><span style=\"font-size: 14pt;\">Borrow from <strong data-start=\"5984\" data-end=\"5995\">Voyager<\/strong>: generate tasks that push frontier skills, mine successful trajectories into a <strong data-start=\"6075\" data-end=\"6095\">callable library<\/strong>, and reuse them compositionally in new worlds\/repos. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2305.16291?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<\/li>\n<li data-start=\"6190\" data-end=\"6420\">\n<h3 data-start=\"6193\" data-end=\"6232\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"6193\" data-end=\"6230\">RLAIF for cheap, scalable rewards<\/strong><\/span><\/h3>\n<ul data-start=\"6236\" data-end=\"6420\">\n<li data-start=\"6236\" data-end=\"6420\">\n<p data-start=\"6238\" data-end=\"6420\"><span style=\"font-size: 14pt;\">Use a strong \u201cteacher LLM\u201d as a <strong data-start=\"6270\" data-end=\"6286\">reward model<\/strong> for preference signals when humans are scarce; fine-tune the policy via RL or direct methods. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2309.00267?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<\/li>\n<li data-start=\"6422\" data-end=\"6451\">\n<h3 data-start=\"6426\" data-end=\"6451\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"6426\" data-end=\"6449\">Risk-aware autonomy<\/strong><\/span><\/h3>\n<\/li>\n<\/ol>\n<ul data-start=\"6455\" data-end=\"6650\">\n<li data-start=\"6455\" data-end=\"6650\">\n<p data-start=\"6457\" data-end=\"6650\"><span style=\"font-size: 14pt;\">Route \u201cdangerous\u201d actions (deleting data, financial trades, PII exposure) through <strong data-start=\"6539\" data-end=\"6570\">mandatory human checkpoints<\/strong> in the graph (HITL edges in LangGraph). <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/www.langchain.com\/langgraph?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">LangChain<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<h2 data-start=\"6657\" data-end=\"6728\"><span style=\"font-size: 18pt; color: #3366ff;\"><strong>4) The RL wave: from RLHF to AI-scaled feedback to pure-RL reasoning<\/strong><\/span><\/h2>\n<ul data-start=\"6730\" data-end=\"7455\">\n<li data-start=\"6730\" data-end=\"7131\">\n<p data-start=\"6732\" data-end=\"7131\"><span style=\"font-size: 14pt;\"><strong data-start=\"6732\" data-end=\"6741\">RLAIF<\/strong> shows that AI-generated preference labels can <strong data-start=\"6788\" data-end=\"6797\">match<\/strong> human-labeled RLHF for alignment, lowering cost and speeding iteration; <strong data-start=\"6870\" data-end=\"6886\">direct-RLAIF<\/strong> bypasses training a separate reward model altogether. Practical takeaway: bootstrap with RLAIF to prototype the reward pipeline, then selectively replace with human gold labels where quality matters most. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2309.00267?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"7133\" data-end=\"7455\">\n<p data-start=\"7135\" data-end=\"7455\"><span style=\"font-size: 14pt;\"><strong data-start=\"7135\" data-end=\"7150\">DeepSeek-R1<\/strong> demonstrates <strong data-start=\"7164\" data-end=\"7195\">pure reinforcement learning<\/strong> can elicit strong <strong data-start=\"7214\" data-end=\"7227\">reasoning<\/strong> without supervised fine-tuning, using reward schemes and curricula that encourage step-by-step thought. Expect more \u201cRL-first\u201d agent stacks\u2014especially for tool-use and long-horizon tasks. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/pdf\/2501.12948?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><span class=\"flex h-4 w-full items-center justify-between absolute\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<p data-start=\"7457\" data-end=\"7661\"><span style=\"font-size: 14pt;\"><strong data-start=\"7457\" data-end=\"7490\">Why this matters to builders:<\/strong> you can now <strong data-start=\"7503\" data-end=\"7521\">train behavior<\/strong> (planning, tool-selection, self-verification) with far fewer human labels\u2014making it realistic to tailor agents to your product\u2019s workflows.<\/span><\/p>\n<h2 data-start=\"7668\" data-end=\"7734\"><span style=\"font-size: 18pt; color: #3366ff;\"><strong>5) Multi-agent collaboration: beyond a single \u201cdo-it-all\u201d model<\/strong><\/span><\/h2>\n<h3 data-start=\"7736\" data-end=\"7776\"><span style=\"color: #800000;\"><strong><span style=\"font-size: 14pt;\">5.1 Collaboration patterns that work<\/span><\/strong><\/span><\/h3>\n<ul data-start=\"7777\" data-end=\"8416\">\n<li data-start=\"7777\" data-end=\"7961\">\n<p data-start=\"7779\" data-end=\"7961\"><span style=\"font-size: 14pt;\"><strong data-start=\"7779\" data-end=\"7815\">Role specialization with AutoGen<\/strong> (Planner, Coder, Tester, Safety) orchestrated through conversational protocols and shared memory\/tools. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2308.08155?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"7962\" data-end=\"8178\">\n<p data-start=\"7964\" data-end=\"8178\"><span style=\"font-size: 14pt;\"><strong data-start=\"7964\" data-end=\"7989\">Debate \/ adjudication<\/strong>: two solvers propose; a <strong data-start=\"8014\" data-end=\"8023\">judge<\/strong> evaluates arguments, enforces rules, and chooses a winner\u2014especially effective on math, code, and fact-checking. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2305.14325?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"8179\" data-end=\"8318\">\n<p data-start=\"8181\" data-end=\"8318\"><span style=\"font-size: 14pt;\"><strong data-start=\"8181\" data-end=\"8203\">Blackboard systems<\/strong>: a shared task board where agents post partial results and pick up subtasks (great for ETL\/reporting pipelines).<\/span><\/p>\n<\/li>\n<li data-start=\"8319\" data-end=\"8416\">\n<p data-start=\"8321\" data-end=\"8416\"><span style=\"font-size: 14pt;\"><strong data-start=\"8321\" data-end=\"8354\">Market\/auction (contract-net)<\/strong>: agents bid on subtasks; useful when capabilities\/costs vary.<\/span><\/p>\n<\/li>\n<\/ul>\n<h3 data-start=\"8418\" data-end=\"8466\"><span style=\"color: #800000;\"><strong><span style=\"font-size: 14pt;\">5.2 Real-world grounding for software agents<\/span><\/strong><\/span><\/h3>\n<ul data-start=\"8467\" data-end=\"8915\">\n<li data-start=\"8467\" data-end=\"8711\">\n<p data-start=\"8469\" data-end=\"8711\"><span style=\"font-size: 14pt;\"><strong data-start=\"8469\" data-end=\"8482\">SWE-bench<\/strong> and <strong data-start=\"8487\" data-end=\"8516\">Agent-Computer Interfaces<\/strong> (ACI) reset the bar by testing agents in <strong data-start=\"8558\" data-end=\"8572\">real repos<\/strong> under tests and CI. Build your evaluation harness in this style\u2014even if you\u2019re not doing coding. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/www.swebench.com\/?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">SWE-Bench<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"8712\" data-end=\"8915\">\n<p data-start=\"8714\" data-end=\"8915\"><span style=\"font-size: 14pt;\">Platforms like <strong data-start=\"8729\" data-end=\"8742\">OpenDevin<\/strong> show how to glue together browser, shell, editor, tests, and coordination across multiple agents\u2014useful as a reference architecture. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/huggingface.co\/papers\/2407.16741?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">Hugging Face<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<h2 data-start=\"8922\" data-end=\"8972\"><span style=\"font-size: 18pt; color: #3366ff;\"><strong>6) Evaluation and operations (MLOps for agents)<\/strong><\/span><\/h2>\n<ol data-start=\"8974\" data-end=\"9594\">\n<li data-start=\"8974\" data-end=\"9164\">\n<p data-start=\"8977\" data-end=\"9164\"><span style=\"font-size: 14pt;\"><strong data-start=\"8977\" data-end=\"8996\">Task-level KPIs<\/strong>: success rate, autonomy ratio (actions per human intervention), cost\/latency, rollback rate, and <em data-start=\"9094\" data-end=\"9117\">post-deployment drift<\/em> (does performance degrade as tasks change?).<\/span><\/p>\n<\/li>\n<li data-start=\"9165\" data-end=\"9267\">\n<p data-start=\"9168\" data-end=\"9267\"><span style=\"font-size: 14pt;\"><strong data-start=\"9168\" data-end=\"9193\">Unit tests for skills<\/strong>: every macro-skill should ship with fixtures, assertions, and timeouts.<\/span><\/p>\n<\/li>\n<li data-start=\"9268\" data-end=\"9356\">\n<p data-start=\"9271\" data-end=\"9356\"><span style=\"font-size: 14pt;\"><strong data-start=\"9271\" data-end=\"9286\">Canary runs<\/strong>: promote new reasoning prompts\/reward models behind traffic splits.<\/span><\/p>\n<\/li>\n<li data-start=\"9357\" data-end=\"9451\">\n<p data-start=\"9360\" data-end=\"9451\"><span style=\"font-size: 14pt;\"><strong data-start=\"9360\" data-end=\"9380\">Replay &amp; diffing<\/strong>: store full state graphs (not just chat logs) to reproduce failures.<\/span><\/p>\n<\/li>\n<li data-start=\"9452\" data-end=\"9594\">\n<p data-start=\"9455\" data-end=\"9594\"><span style=\"font-size: 14pt;\"><strong data-start=\"9455\" data-end=\"9469\">Benchmarks<\/strong>: adopt a SWE-bench-style \u201clive\u201d set for your domain (weekly refresh of real issues). <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/swe-bench-live.github.io\/?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">SWE-bench Live<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ol>\n<h2 data-start=\"9601\" data-end=\"9649\"><span style=\"font-size: 18pt; color: #3366ff;\"><strong>7) Safety &amp; governance you can operationalize<\/strong><\/span><\/h2>\n<ul data-start=\"9651\" data-end=\"10259\">\n<li data-start=\"9651\" data-end=\"9849\">\n<p data-start=\"9653\" data-end=\"9849\"><span style=\"font-size: 14pt;\"><strong data-start=\"9653\" data-end=\"9676\">Sandboxed execution<\/strong> (containers, read-write allowlists), <strong data-start=\"9714\" data-end=\"9740\">capability permissions<\/strong> (per-tool scopes), and <strong data-start=\"9764\" data-end=\"9782\">mandatory HITL<\/strong> on irreversible actions. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/www.langchain.com\/langgraph?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">LangChain<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"9850\" data-end=\"9931\">\n<p data-start=\"9852\" data-end=\"9931\"><span style=\"font-size: 14pt;\"><strong data-start=\"9852\" data-end=\"9881\">Content and data controls<\/strong> at graph edges (PII detection, policy filters).<\/span><\/p>\n<\/li>\n<li data-start=\"9932\" data-end=\"10052\">\n<p data-start=\"9934\" data-end=\"10052\"><span style=\"font-size: 14pt;\"><strong data-start=\"9934\" data-end=\"9950\">Auditability<\/strong>: persist action logs + thought summaries (not raw private chain-of-thought) for compliance reviews.<\/span><\/p>\n<\/li>\n<li data-start=\"10053\" data-end=\"10259\">\n<p data-start=\"10055\" data-end=\"10259\"><span style=\"font-size: 14pt;\"><strong data-start=\"10055\" data-end=\"10071\">Risk posture<\/strong>: as autonomy and reasoning scale (e.g., via DeepSeek-style RL), re-assess threat models\u2014leading researchers have raised correlated safety concerns. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/www.theguardian.com\/technology\/2025\/jan\/29\/deepseek-artificial-intelligence-ai-safety-risk-yoshua-bengio?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">The Guardian<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<h2 data-start=\"10266\" data-end=\"10306\"><span style=\"font-size: 18pt; color: #3366ff;\"><strong>8) Implementation blueprint (90 days)<\/strong><\/span><\/h2>\n<h3 data-start=\"10308\" data-end=\"10340\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"10308\" data-end=\"10340\">Days 0\u201315: Skeleton &amp; safety<\/strong><\/span><\/h3>\n<ul data-start=\"10341\" data-end=\"10548\">\n<li data-start=\"10341\" data-end=\"10432\">\n<p data-start=\"10343\" data-end=\"10432\"><span style=\"font-size: 14pt;\">Choose runtime (<strong data-start=\"10359\" data-end=\"10372\">LangGraph<\/strong> or equivalent) and memory layer (<strong data-start=\"10406\" data-end=\"10422\">MemGPT-style<\/strong> tiers).<\/span><\/p>\n<\/li>\n<li data-start=\"10433\" data-end=\"10548\">\n<p data-start=\"10435\" data-end=\"10548\"><span style=\"font-size: 14pt;\">Implement ReAct loop + tool adapters; add HITL gates for sensitive edges. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2210.03629?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<h3 data-start=\"10550\" data-end=\"10586\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"10550\" data-end=\"10586\">Days 16\u201345: Reliability &amp; skills<\/strong><\/span><\/h3>\n<ul data-start=\"10587\" data-end=\"10796\">\n<li data-start=\"10587\" data-end=\"10652\">\n<p data-start=\"10589\" data-end=\"10652\"><span style=\"font-size: 14pt;\">Add <strong data-start=\"10593\" data-end=\"10613\">self-consistency<\/strong> decoding; wire <strong data-start=\"10629\" data-end=\"10642\">Reflexion<\/strong> memory.<\/span><\/p>\n<\/li>\n<li data-start=\"10653\" data-end=\"10796\">\n<p data-start=\"10655\" data-end=\"10796\"><span style=\"font-size: 14pt;\">Extract recurring steps into <strong data-start=\"10684\" data-end=\"10694\">skills<\/strong> with unit tests; layer <strong data-start=\"10718\" data-end=\"10740\">lookahead decoding<\/strong> to cut latency. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2203.11171?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><span class=\"flex h-4 w-full items-center justify-between absolute\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<h3 data-start=\"10798\" data-end=\"10832\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"10798\" data-end=\"10832\">Days 46\u201375: Collaboration &amp; RL<\/strong><\/span><\/h3>\n<ul data-start=\"10833\" data-end=\"11078\">\n<li data-start=\"10833\" data-end=\"10895\">\n<p data-start=\"10835\" data-end=\"10895\"><span style=\"font-size: 14pt;\">Introduce <strong data-start=\"10845\" data-end=\"10856\">AutoGen<\/strong> roles (planner\/coder\/tester\/safety).<\/span><\/p>\n<\/li>\n<li data-start=\"10896\" data-end=\"11078\">\n<p data-start=\"10898\" data-end=\"11078\"><span style=\"font-size: 14pt;\">Stand up <strong data-start=\"10907\" data-end=\"10916\">RLAIF<\/strong> for scalable preference rewards; if feasible, pilot <strong data-start=\"10969\" data-end=\"10980\">pure-RL<\/strong> fine-tuning on a narrow task with simulated environments. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/microsoft.github.io\/autogen\/0.2\/docs\/Use-Cases\/agent_chat\/?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">Microsoft GitHub<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><span class=\"flex h-4 w-full items-center justify-between absolute\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+2<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<h3 data-start=\"11080\" data-end=\"11118\"><span style=\"font-size: 14pt; color: #800000;\"><strong data-start=\"11080\" data-end=\"11118\">Days 76\u201390: Evaluation &amp; hardening<\/strong><\/span><\/h3>\n<ul data-start=\"11119\" data-end=\"11320\">\n<li data-start=\"11119\" data-end=\"11320\">\n<p data-start=\"11121\" data-end=\"11320\"><span style=\"font-size: 14pt;\">Build a <strong data-start=\"11129\" data-end=\"11147\">live benchmark<\/strong> mirroring SWE-bench\u2019s philosophy for your domain; add ACIs for structured actions; ship dashboards for autonomy, cost, and rollback. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/www.swebench.com\/?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">SWE-Bench<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<h2 data-start=\"11327\" data-end=\"11387\"><span style=\"font-size: 18pt; color: #3366ff;\"><strong>Appendix A \u2014 Quick reference to key research &amp; frameworks<\/strong><\/span><\/h2>\n<ul data-start=\"11389\" data-end=\"12837\">\n<li data-start=\"11389\" data-end=\"11532\">\n<p data-start=\"11391\" data-end=\"11532\"><span style=\"font-size: 14pt;\"><strong data-start=\"11391\" data-end=\"11400\">ReAct<\/strong>: Integrates reasoning traces with actions to reduce hallucinations and handle exceptions. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2210.03629?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"11533\" data-end=\"11687\">\n<p data-start=\"11535\" data-end=\"11687\"><span style=\"font-size: 14pt;\"><strong data-start=\"11535\" data-end=\"11575\">Tree-of-Thoughts \/ Graph-of-Thoughts<\/strong>: Structured search over reasoning improves outcomes and cost control. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2305.10601?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"11688\" data-end=\"11813\">\n<p data-start=\"11690\" data-end=\"11813\"><span style=\"font-size: 14pt;\"><strong data-start=\"11690\" data-end=\"11710\">Self-Consistency<\/strong>: Sample multiple reasoning paths; pick the consensus answer. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2203.11171?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"11814\" data-end=\"11960\">\n<p data-start=\"11816\" data-end=\"11960\"><span style=\"font-size: 14pt;\"><strong data-start=\"11816\" data-end=\"11829\">Reflexion<\/strong>: Verbal self-critique + episodic memory that improves next tries without weight updates. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2303.11366?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"11961\" data-end=\"12068\">\n<p data-start=\"11963\" data-end=\"12068\"><span style=\"font-size: 14pt;\"><strong data-start=\"11963\" data-end=\"11971\">LATS<\/strong>: Unifies planning, acting, reasoning with tree search. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2310.04406?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"12069\" data-end=\"12180\">\n<p data-start=\"12071\" data-end=\"12180\"><span style=\"font-size: 14pt;\"><strong data-start=\"12071\" data-end=\"12081\">MemGPT<\/strong>: Tiered memory + interrupts; treat the agent like an OS. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2310.08560?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"12181\" data-end=\"12303\">\n<p data-start=\"12183\" data-end=\"12303\"><span style=\"font-size: 14pt;\"><strong data-start=\"12183\" data-end=\"12196\">LangGraph<\/strong>: Stateful, cyclical agent graphs with HITL and moderation hooks. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/www.langchain.com\/langgraph?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">LangChain<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"12304\" data-end=\"12418\">\n<p data-start=\"12306\" data-end=\"12418\"><span style=\"font-size: 14pt;\"><strong data-start=\"12306\" data-end=\"12317\">AutoGen<\/strong>: Orchestrate multi-agent conversations, tools, and humans. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2308.08155?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"12419\" data-end=\"12547\">\n<p data-start=\"12421\" data-end=\"12547\"><span style=\"font-size: 14pt;\"><strong data-start=\"12421\" data-end=\"12430\">RLAIF<\/strong>: Replace human labels with AI preference signals; includes <em data-start=\"12490\" data-end=\"12504\">direct-RLAIF<\/em>. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/abs\/2309.00267?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"12548\" data-end=\"12693\">\n<p data-start=\"12550\" data-end=\"12693\"><span style=\"font-size: 14pt;\"><strong data-start=\"12550\" data-end=\"12565\">DeepSeek-R1<\/strong>: Pure-RL training elicits reasoning\u2014pointing to lower-label, higher-autonomy futures. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/arxiv.org\/pdf\/2501.12948?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between overflow-hidden\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">arXiv<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<li data-start=\"12694\" data-end=\"12837\">\n<p data-start=\"12696\" data-end=\"12837\"><span style=\"font-size: 14pt;\"><strong data-start=\"12696\" data-end=\"12715\">SWE-bench + ACI<\/strong>: Real-world coding tasks; structured interfaces materially improve agent success. <span class=\"\" data-state=\"closed\"><span class=\"ms-1 inline-flex max-w-full items-center relative top-[-0.094rem] animate-[show_150ms_ease-in]\" data-testid=\"webpage-citation-pill\"><a class=\"flex h-4.5 overflow-hidden rounded-xl px-2 text-[9px] font-medium text-token-text-secondary! bg-[#F4F4F4]! dark:bg-[#303030]! transition-colors duration-150 ease-in-out\" href=\"https:\/\/www.swebench.com\/?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noopener\"><span class=\"relative start-0 bottom-0 flex h-full w-full items-center\"><span class=\"flex h-4 w-full items-center justify-between\"><span class=\"max-w-[15ch] grow truncate overflow-hidden text-center\">SWE-Bench<\/span><span class=\"-me-1 flex h-full items-center rounded-full px-1 text-[#8F8F8F]\">+1<\/span><\/span><\/span><\/a><\/span><\/span><\/span><\/p>\n<\/li>\n<\/ul>\n<h2 data-start=\"12844\" data-end=\"12885\"><span style=\"font-size: 18pt; color: #3366ff;\"><strong>Appendix B \u2014 Minimal starter checklist<\/strong><\/span><\/h2>\n<ul class=\"contains-task-list\" data-start=\"12887\" data-end=\"13410\">\n<li class=\"task-list-item\" data-start=\"12887\" data-end=\"12932\">\n<p data-start=\"12893\" data-end=\"12932\"><span style=\"font-size: 14pt;\"><input disabled=\"disabled\" type=\"checkbox\" \/> ReAct loop with explicit state object<\/span><\/p>\n<\/li>\n<li class=\"task-list-item\" data-start=\"12933\" data-end=\"13028\">\n<p data-start=\"12939\" data-end=\"13028\"><span style=\"font-size: 14pt;\"><input disabled=\"disabled\" type=\"checkbox\" \/> Thought search: start with <strong data-start=\"12966\" data-end=\"12986\">self-consistency<\/strong>, add small <strong data-start=\"12998\" data-end=\"13005\">ToT<\/strong> depth for hard tasks<\/span><\/p>\n<\/li>\n<li class=\"task-list-item\" data-start=\"13029\" data-end=\"13100\">\n<p data-start=\"13035\" data-end=\"13100\"><span style=\"font-size: 14pt;\"><input disabled=\"disabled\" type=\"checkbox\" \/> Memory tiers (scratchpad \u2192 episodic \u2192 semantic \u2192 skill library)<\/span><\/p>\n<\/li>\n<li class=\"task-list-item\" data-start=\"13101\" data-end=\"13153\">\n<p data-start=\"13103\" data-end=\"13153\"><span style=\"font-size: 14pt;\"><input disabled=\"disabled\" type=\"checkbox\" \/> <strong data-start=\"13107\" data-end=\"13120\">Validator<\/strong> for every tool and macro-skill<\/span><\/p>\n<\/li>\n<li class=\"task-list-item\" data-start=\"13154\" data-end=\"13219\">\n<p data-start=\"13160\" data-end=\"13219\"><span style=\"font-size: 14pt;\"><input disabled=\"disabled\" type=\"checkbox\" \/> Conditional <strong data-start=\"13172\" data-end=\"13190\">debate + judge<\/strong> only on high-entropy tasks<\/span><\/p>\n<\/li>\n<li class=\"task-list-item\" data-start=\"13220\" data-end=\"13267\">\n<p data-start=\"13222\" data-end=\"13267\"><span style=\"font-size: 14pt;\"><input disabled=\"disabled\" type=\"checkbox\" \/> <strong data-start=\"13226\" data-end=\"13239\">Lookahead<\/strong> decoding for long outputs<\/span><\/p>\n<\/li>\n<li class=\"task-list-item\" data-start=\"13268\" data-end=\"13334\">\n<p data-start=\"13270\" data-end=\"13334\"><span style=\"font-size: 14pt;\"><input disabled=\"disabled\" type=\"checkbox\" \/> <strong data-start=\"13274\" data-end=\"13283\">RLAIF<\/strong> pipeline for scalable rewards; log reward traces<\/span><\/p>\n<\/li>\n<li class=\"task-list-item\" data-start=\"13335\" data-end=\"13410\">\n<p data-start=\"13337\" data-end=\"13410\"><span style=\"font-size: 14pt;\"><input disabled=\"disabled\" type=\"checkbox\" \/> <strong data-start=\"13341\" data-end=\"13359\">Live benchmark<\/strong> and <strong data-start=\"13364\" data-end=\"13371\">ACI<\/strong> harness; ship autonomy\/cost dashboards<\/span><\/p>\n<\/li>\n<\/ul>\n<h3 data-start=\"13417\" data-end=\"13431\"><span style=\"color: #800000;\"><strong><span style=\"font-size: 14pt;\">Final note<\/span><\/strong><\/span><\/h3>\n<p data-start=\"13433\" data-end=\"13700\"><span style=\"font-size: 14pt;\">The \u201cagentic frontier\u201d is moving fast, but the shape is clear: <strong data-start=\"13496\" data-end=\"13598\">structured reasoning, explicit state, skill libraries, scalable rewards, and multi-agent protocols<\/strong>. <\/span><\/p>\n<p data-start=\"13433\" data-end=\"13700\"><span style=\"font-size: 14pt;\">Start with the simple hacks above; then layer in RL and collaboration where they buy you clear wins.<\/span><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Peek behind the curtain to reveal the hidden dynamics of developing agentic AI systems\u2014plus hands-on \u201chacks\u201d to boost autonomy and efficiency with the latest advances in reinforcement learning and multi-agent collaboration. Executive Summary Agentic AI \u2260 \u201ca chatbot with tools.\u201d Truly agentic systems can decide, act, learn, and coordinate\u2014under uncertainty, across long horizons, with measurable outcomes and guardrails. The current best practice blends: Reasoning + acting loops (plan, act, observe, reflect) with tool-use. arXiv+1 Search over thoughts (trees\/graphs), not just single chains. arXiv+2NeurIPS Proceedings+2 Purpose-built memory and stateful runtimes (episodic\/semantic\/procedural state). arXiv+1 Reinforcement learning that increasingly reduces reliance on supervised&#8230; <\/p>\n","protected":false},"author":1,"featured_media":904,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[108],"tags":[119],"class_list":["post-902","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-agentic-ai","tag-building-agentic-ai"],"aioseo_notices":[],"aioseo_head":"\n\t\t<!-- All in One SEO 5.0.4 - aioseo.com -->\n\t<meta name=\"description\" content=\"Peek behind the curtain to reveal the hidden dynamics of developing agentic AI systems\u2014plus hands-on \u201chacks\u201d to boost autonomy and efficiency with the latest advances in reinforcement learning and multi-agent collaboration. Executive Summary Agentic AI \u2260 \u201ca chatbot with tools.\u201d Truly agentic systems can decide, act, learn, and coordinate\u2014under uncertainty, across long horizons, with measurable\" \/>\n\t<meta name=\"robots\" content=\"max-image-preview:large\" \/>\n\t<meta name=\"author\" content=\"aihatch.online\"\/>\n\t<link rel=\"canonical\" href=\"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/\" \/>\n\t\t<meta name=\"generator\" content=\"All in One SEO (AIOSEO) 5.0.4\" \/>\n\t\t<meta property=\"og:locale\" content=\"en_GB\" \/>\n\t\t<meta property=\"og:site_name\" content=\"aihatch.online - Discover the Power of AI\" \/>\n\t\t<meta property=\"og:type\" content=\"article\" \/>\n\t\t<meta property=\"og:title\" content=\"Building Truly Agentic AI Systems - aihatch.online\" \/>\n\t\t<meta property=\"og:description\" content=\"Peek behind the curtain to reveal the hidden dynamics of developing agentic AI systems\u2014plus hands-on \u201chacks\u201d to boost autonomy and efficiency with the latest advances in reinforcement learning and multi-agent collaboration. Executive Summary Agentic AI \u2260 \u201ca chatbot with tools.\u201d Truly agentic systems can decide, act, learn, and coordinate\u2014under uncertainty, across long horizons, with measurable\" \/>\n\t\t<meta property=\"og:url\" content=\"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/\" \/>\n\t\t<meta property=\"article:published_time\" content=\"2025-09-14T10:30:14+00:00\" \/>\n\t\t<meta property=\"article:modified_time\" content=\"2025-09-14T10:38:50+00:00\" \/>\n\t\t<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n\t\t<meta name=\"twitter:title\" content=\"Building Truly Agentic AI Systems - aihatch.online\" \/>\n\t\t<meta name=\"twitter:description\" content=\"Peek behind the curtain to reveal the hidden dynamics of developing agentic AI systems\u2014plus hands-on \u201chacks\u201d to boost autonomy and efficiency with the latest advances in reinforcement learning and multi-agent collaboration. Executive Summary Agentic AI \u2260 \u201ca chatbot with tools.\u201d Truly agentic systems can decide, act, learn, and coordinate\u2014under uncertainty, across long horizons, with measurable\" \/>\n\t\t<script type=\"application\/ld+json\" class=\"aioseo-schema\">\n\t\t\t{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"BlogPosting\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/building-truly-agentic-ai-systems-2\\\/#blogposting\",\"name\":\"Building Truly Agentic AI Systems - aihatch.online\",\"headline\":\"Building Truly Agentic AI Systems\",\"author\":{\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/author\\\/aihatch-online\\\/#author\"},\"publisher\":{\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/#organization\"},\"image\":{\"@type\":\"ImageObject\",\"url\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/wp-content\\\/uploads\\\/2025\\\/09\\\/AgenticAI.png\",\"width\":1024,\"height\":1024,\"caption\":\"Building Truly Agentic AI Systems\"},\"datePublished\":\"2025-09-14T06:30:14-04:00\",\"dateModified\":\"2025-09-14T06:38:50-04:00\",\"inLanguage\":\"en-GB\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/building-truly-agentic-ai-systems-2\\\/#webpage\"},\"isPartOf\":{\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/building-truly-agentic-ai-systems-2\\\/#webpage\"},\"articleSection\":\"Agentic AI, Building Agentic AI\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/building-truly-agentic-ai-systems-2\\\/#breadcrumblist\",\"itemListElement\":[{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/#listItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/\",\"nextItem\":{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/category\\\/agentic-ai\\\/#listItem\",\"name\":\"Agentic AI\"}},{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/category\\\/agentic-ai\\\/#listItem\",\"position\":2,\"name\":\"Agentic AI\",\"item\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/category\\\/agentic-ai\\\/\",\"nextItem\":{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/building-truly-agentic-ai-systems-2\\\/#listItem\",\"name\":\"Building Truly Agentic AI Systems\"},\"previousItem\":{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/#listItem\",\"name\":\"Home\"}},{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/building-truly-agentic-ai-systems-2\\\/#listItem\",\"position\":3,\"name\":\"Building Truly Agentic AI Systems\",\"previousItem\":{\"@type\":\"ListItem\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/category\\\/agentic-ai\\\/#listItem\",\"name\":\"Agentic AI\"}}]},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/#organization\",\"name\":\"aihatch.online\",\"description\":\"Discover the Power of AI\",\"url\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/author\\\/aihatch-online\\\/#author\",\"url\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/author\\\/aihatch-online\\\/\",\"name\":\"aihatch.online\",\"image\":{\"@type\":\"ImageObject\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/building-truly-agentic-ai-systems-2\\\/#authorImage\",\"url\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/wp-content\\\/litespeed\\\/avatar\\\/d1ce21c9df09238dba12698bcf6a5a0e.jpg?ver=1793114063\",\"width\":96,\"height\":96,\"caption\":\"aihatch.online\"}},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/building-truly-agentic-ai-systems-2\\\/#webpage\",\"url\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/building-truly-agentic-ai-systems-2\\\/\",\"name\":\"Building Truly Agentic AI Systems - aihatch.online\",\"description\":\"Peek behind the curtain to reveal the hidden dynamics of developing agentic AI systems\\u2014plus hands-on \\u201chacks\\u201d to boost autonomy and efficiency with the latest advances in reinforcement learning and multi-agent collaboration. Executive Summary Agentic AI \\u2260 \\u201ca chatbot with tools.\\u201d Truly agentic systems can decide, act, learn, and coordinate\\u2014under uncertainty, across long horizons, with measurable\",\"inLanguage\":\"en-GB\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/#website\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/building-truly-agentic-ai-systems-2\\\/#breadcrumblist\"},\"author\":{\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/author\\\/aihatch-online\\\/#author\"},\"creator\":{\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/author\\\/aihatch-online\\\/#author\"},\"image\":{\"@type\":\"ImageObject\",\"url\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/wp-content\\\/uploads\\\/2025\\\/09\\\/AgenticAI.png\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/building-truly-agentic-ai-systems-2\\\/#mainImage\",\"width\":1024,\"height\":1024,\"caption\":\"Building Truly Agentic AI Systems\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/building-truly-agentic-ai-systems-2\\\/#mainImage\"},\"datePublished\":\"2025-09-14T06:30:14-04:00\",\"dateModified\":\"2025-09-14T06:38:50-04:00\"},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/#website\",\"url\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/\",\"name\":\"aihatch.online\",\"description\":\"Discover the Power of AI\",\"inLanguage\":\"en-GB\",\"publisher\":{\"@id\":\"https:\\\/\\\/infohatch.net\\\/ai\\\/#organization\"}}]}\n\t\t<\/script>\n\t\t<!-- All in One SEO -->\n\n","aioseo_head_json":{"title":"Building Truly Agentic AI Systems - aihatch.online","description":"Peek behind the curtain to reveal the hidden dynamics of developing agentic AI systems\u2014plus hands-on \u201chacks\u201d to boost autonomy and efficiency with the latest advances in reinforcement learning and multi-agent collaboration. Executive Summary Agentic AI \u2260 \u201ca chatbot with tools.\u201d Truly agentic systems can decide, act, learn, and coordinate\u2014under uncertainty, across long horizons, with measurable","canonical_url":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/","robots":"max-image-preview:large","keywords":"","webmasterTools":{"miscellaneous":""},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"BlogPosting","@id":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/#blogposting","name":"Building Truly Agentic AI Systems - aihatch.online","headline":"Building Truly Agentic AI Systems","author":{"@id":"https:\/\/infohatch.net\/ai\/author\/aihatch-online\/#author"},"publisher":{"@id":"https:\/\/infohatch.net\/ai\/#organization"},"image":{"@type":"ImageObject","url":"https:\/\/infohatch.net\/ai\/wp-content\/uploads\/2025\/09\/AgenticAI.png","width":1024,"height":1024,"caption":"Building Truly Agentic AI Systems"},"datePublished":"2025-09-14T06:30:14-04:00","dateModified":"2025-09-14T06:38:50-04:00","inLanguage":"en-GB","mainEntityOfPage":{"@id":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/#webpage"},"isPartOf":{"@id":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/#webpage"},"articleSection":"Agentic AI, Building Agentic AI"},{"@type":"BreadcrumbList","@id":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/#breadcrumblist","itemListElement":[{"@type":"ListItem","@id":"https:\/\/infohatch.net\/ai\/#listItem","position":1,"name":"Home","item":"https:\/\/infohatch.net\/ai\/","nextItem":{"@type":"ListItem","@id":"https:\/\/infohatch.net\/ai\/category\/agentic-ai\/#listItem","name":"Agentic AI"}},{"@type":"ListItem","@id":"https:\/\/infohatch.net\/ai\/category\/agentic-ai\/#listItem","position":2,"name":"Agentic AI","item":"https:\/\/infohatch.net\/ai\/category\/agentic-ai\/","nextItem":{"@type":"ListItem","@id":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/#listItem","name":"Building Truly Agentic AI Systems"},"previousItem":{"@type":"ListItem","@id":"https:\/\/infohatch.net\/ai\/#listItem","name":"Home"}},{"@type":"ListItem","@id":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/#listItem","position":3,"name":"Building Truly Agentic AI Systems","previousItem":{"@type":"ListItem","@id":"https:\/\/infohatch.net\/ai\/category\/agentic-ai\/#listItem","name":"Agentic AI"}}]},{"@type":"Organization","@id":"https:\/\/infohatch.net\/ai\/#organization","name":"aihatch.online","description":"Discover the Power of AI","url":"https:\/\/infohatch.net\/ai\/"},{"@type":"Person","@id":"https:\/\/infohatch.net\/ai\/author\/aihatch-online\/#author","url":"https:\/\/infohatch.net\/ai\/author\/aihatch-online\/","name":"aihatch.online","image":{"@type":"ImageObject","@id":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/#authorImage","url":"https:\/\/infohatch.net\/ai\/wp-content\/litespeed\/avatar\/d1ce21c9df09238dba12698bcf6a5a0e.jpg?ver=1793114063","width":96,"height":96,"caption":"aihatch.online"}},{"@type":"WebPage","@id":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/#webpage","url":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/","name":"Building Truly Agentic AI Systems - aihatch.online","description":"Peek behind the curtain to reveal the hidden dynamics of developing agentic AI systems\u2014plus hands-on \u201chacks\u201d to boost autonomy and efficiency with the latest advances in reinforcement learning and multi-agent collaboration. Executive Summary Agentic AI \u2260 \u201ca chatbot with tools.\u201d Truly agentic systems can decide, act, learn, and coordinate\u2014under uncertainty, across long horizons, with measurable","inLanguage":"en-GB","isPartOf":{"@id":"https:\/\/infohatch.net\/ai\/#website"},"breadcrumb":{"@id":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/#breadcrumblist"},"author":{"@id":"https:\/\/infohatch.net\/ai\/author\/aihatch-online\/#author"},"creator":{"@id":"https:\/\/infohatch.net\/ai\/author\/aihatch-online\/#author"},"image":{"@type":"ImageObject","url":"https:\/\/infohatch.net\/ai\/wp-content\/uploads\/2025\/09\/AgenticAI.png","@id":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/#mainImage","width":1024,"height":1024,"caption":"Building Truly Agentic AI Systems"},"primaryImageOfPage":{"@id":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/#mainImage"},"datePublished":"2025-09-14T06:30:14-04:00","dateModified":"2025-09-14T06:38:50-04:00"},{"@type":"WebSite","@id":"https:\/\/infohatch.net\/ai\/#website","url":"https:\/\/infohatch.net\/ai\/","name":"aihatch.online","description":"Discover the Power of AI","inLanguage":"en-GB","publisher":{"@id":"https:\/\/infohatch.net\/ai\/#organization"}}]},"og:locale":"en_GB","og:site_name":"aihatch.online - Discover the Power of AI","og:type":"article","og:title":"Building Truly Agentic AI Systems - aihatch.online","og:description":"Peek behind the curtain to reveal the hidden dynamics of developing agentic AI systems\u2014plus hands-on \u201chacks\u201d to boost autonomy and efficiency with the latest advances in reinforcement learning and multi-agent collaboration. Executive Summary Agentic AI \u2260 \u201ca chatbot with tools.\u201d Truly agentic systems can decide, act, learn, and coordinate\u2014under uncertainty, across long horizons, with measurable","og:url":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/","article:published_time":"2025-09-14T10:30:14+00:00","article:modified_time":"2025-09-14T10:38:50+00:00","twitter:card":"summary_large_image","twitter:title":"Building Truly Agentic AI Systems - aihatch.online","twitter:description":"Peek behind the curtain to reveal the hidden dynamics of developing agentic AI systems\u2014plus hands-on \u201chacks\u201d to boost autonomy and efficiency with the latest advances in reinforcement learning and multi-agent collaboration. Executive Summary Agentic AI \u2260 \u201ca chatbot with tools.\u201d Truly agentic systems can decide, act, learn, and coordinate\u2014under uncertainty, across long horizons, with measurable"},"aioseo_meta_data":{"post_id":"902","title":"#post_title #separator_sa #site_title","description":"#post_excerpt","keywords":null,"keyphrases":{"focus":[],"additional":[]},"primary_term":null,"canonical_url":null,"og_title":null,"og_description":null,"og_object_type":"default","og_image_type":"default","og_image_url":null,"og_image_width":null,"og_image_height":null,"og_image_custom_url":null,"og_image_custom_fields":null,"og_video":"","og_custom_url":null,"og_article_section":null,"og_article_tags":null,"twitter_use_og":false,"twitter_card":"default","twitter_image_type":"default","twitter_image_url":null,"twitter_image_custom_url":null,"twitter_image_custom_fields":null,"twitter_title":null,"twitter_description":null,"schema":{"blockGraphs":[],"customGraphs":[],"default":{"data":{"Article":[],"Course":[],"Dataset":[],"FAQPage":[],"Movie":[],"Person":[],"Product":[],"ProductReview":[],"Car":[],"Recipe":[],"Service":[],"SoftwareApplication":[],"WebPage":[]},"graphName":"BlogPosting","isEnabled":true},"graphs":[]},"schema_type":"default","schema_type_options":null,"pillar_content":false,"robots_default":true,"robots_noindex":false,"robots_noarchive":false,"robots_nosnippet":false,"robots_nofollow":false,"robots_noimageindex":false,"robots_noodp":false,"robots_notranslate":false,"robots_max_snippet":"-1","robots_max_videopreview":"-1","robots_max_imagepreview":"large","priority":null,"frequency":"default","local_seo":null,"breadcrumb_settings":null,"limit_modified_date":false,"created":"2025-09-14 10:30:15","updated":"2026-08-16 13:17:00","ai":null,"seo_analyzer_scan_date":null,"focus_keyword":null,"additional_keywords":null,"truseo_locale":null},"aioseo_breadcrumb":"<div class=\"aioseo-breadcrumbs\"><span class=\"aioseo-breadcrumb\">\n\t\t\t<a href=\"https:\/\/infohatch.net\/ai\/\" title=\"Home\">Home<\/a>\n\t\t<\/span><span class=\"aioseo-breadcrumb-separator\">\u00bb<\/span><span class=\"aioseo-breadcrumb\">\n\t\t\t<a href=\"https:\/\/infohatch.net\/ai\/category\/agentic-ai\/\" title=\"Agentic AI\">Agentic AI<\/a>\n\t\t<\/span><span class=\"aioseo-breadcrumb-separator\">\u00bb<\/span><span class=\"aioseo-breadcrumb\">\n\t\t\tBuilding Truly Agentic AI Systems\n\t\t<\/span><\/div>","aioseo_breadcrumb_json":[{"label":"Home","link":"https:\/\/infohatch.net\/ai\/"},{"label":"Agentic AI","link":"https:\/\/infohatch.net\/ai\/category\/agentic-ai\/"},{"label":"Building Truly Agentic AI Systems","link":"https:\/\/infohatch.net\/ai\/building-truly-agentic-ai-systems-2\/"}],"_links":{"self":[{"href":"https:\/\/infohatch.net\/ai\/wp-json\/wp\/v2\/posts\/902","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infohatch.net\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infohatch.net\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infohatch.net\/ai\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/infohatch.net\/ai\/wp-json\/wp\/v2\/comments?post=902"}],"version-history":[{"count":1,"href":"https:\/\/infohatch.net\/ai\/wp-json\/wp\/v2\/posts\/902\/revisions"}],"predecessor-version":[{"id":903,"href":"https:\/\/infohatch.net\/ai\/wp-json\/wp\/v2\/posts\/902\/revisions\/903"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infohatch.net\/ai\/wp-json\/wp\/v2\/media\/904"}],"wp:attachment":[{"href":"https:\/\/infohatch.net\/ai\/wp-json\/wp\/v2\/media?parent=902"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infohatch.net\/ai\/wp-json\/wp\/v2\/categories?post=902"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infohatch.net\/ai\/wp-json\/wp\/v2\/tags?post=902"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}