BabyAGI and AutoGPT fizzled. What's left of 2023: the agent that thinks in a loop

In short

• ReAct (2022) alternates thought, action and observation, in a loop.
• In the ReAct paper, hallucination accounts for 56% of the chain-of-thought failures examined, 0% of ReAct's, which on the other hand gets lost in its own reasoning more often.
• Reflexion (2023): the agent notes why it failed and uses that on the next try.

March 2023, the rush

Chapter "ReAct, BabyAGI & AutoGPT: the agent boom" (14:00). First came ReAct, in 2022: reason plus act. "So instead of thinking once and then answering the question, what the model does here is actually a loop," Diyi Yang explains. It thinks about what to do, acts, looks at what happened, and goes from there.

Then ChatGPT, on 30/11/2022, then GPT-4. An agent depends on its base model, the professor insists, and GPT-4 follows long instructions and plans more reliably.

Within two weeks, two open-source projects changed how people thought about agents. BabyAGI: you give the model a goal, it writes its own task list, works through it one task at a time and adds new ones as it goes. AutoGPT: GPT-4 hooked up to tools, web search, files, that keeps going on its own. Ideas ahead of their time, she says, that didn't really work, but that grabbed attention and opened up imaginations. That's probably when the word agent entered everyday language.

Will it plateau?

Chapter "Q&A: Will model capabilities saturate?" (19:06). Her answer: impossible to predict, so much happens in two or three months these days. She mentions recursive self-improvement (RSI), when an AI improves itself, with interesting results on open maths problems: "I think with RSI, this pattern will get accelerated a lot, so let's see."

Why do the best results cost more? The price follows the number of tokens processed, and thinking longer generally improves results. Why do new models do better? Many reasons, and, she admits, we don't know much about how they're trained. A Stanford professor saying so in front of her students: that's worth ten charts.

How an agent thinks

Chapter "Planning & reasoning: CoT, ReAct, Reflexion" (28:19). For a language model, reasoning mostly means chain of thought: it writes out its steps before answering. Diyi Yang warns that this text may not be faithful to what actually happens inside the model.

Her example: an agent is cooking, the salt has run out, and it turns right. Why? With its written reasoning, it all becomes clear: it needs a substitute for salt, and it's in the cabinet to the right (soy sauce, on the slide). The same reasoning will work again in similar situations, and "you can see why the agent did what it did".

Then the Apple remote. The question: apart from the Apple Remote, what other device can control the program it was originally designed to interact with? Chain of thought alone answers iPhone, iPad, iPod Touch. Wrong, with flawless-looking reasoning. "This is actually more dangerous, because you have a very convincing reasoning trace that's wrong," she says. ReAct searches Wikipedia, learns that the remote controlled the Front Row software, can't find it, fixes its search, and answers: the keyboard's function keys.

I opened the ReAct paper for the figures the professor puts on screen. Its authors examined by hand 50 failed answers from each method, on those questions. Hallucination makes up 56% of chain of thought's failures, 0% of ReAct's. But ReAct gets lost in its reasoning more often, 47% versus 16%, for example by repeating its previous thoughts and actions in a loop.

Reflexion goes further. The agent makes an attempt, an evaluator says right or wrong, and the agent writes down in plain language what went wrong; that text becomes a memory for the next attempt. According to the professor, it does clearly better than ReAct alone, with fewer hallucinations. Its authors report 91% of programs right on the first try on the HumanEval test, against 80% for GPT-4 alone according to their own table (OpenAI reported 67% in March 2023).

ReAct or chain of thought: questions from the room

Chapter "Q&A: ReAct vs. chain of thought" (36:52). Is Reflexion retraining? No, she answers: none of the model's settings are changed, and most of these self-reflection techniques work through the prompt. The paper talks about "reinforcing" the agent, but it isn't reinforcement learning.

Zero hallucinations but lots of reasoning errors? You can fail in other ways, on an ambiguous expected answer for instance. And the fundamental difference? Chain of thought: "it's literally just like one prompt, one pass." ReAct is a loop: the model can fix its own mistakes or go and get new information to update itself.

My take: BabyAGI and AutoGPT wanted autonomy right away. What stuck is more modest and more solid: think before acting, look at the result, write down your misses. Kira calls it "keeping a notebook". She's not wrong.

Words to remember

• Chain of thought: the model writes out its reasoning steps before answering.
• Hallucination: a fact or a line of reasoning made up by the model.
• Token: a piece of a word, the unit the model reads and writes, and that you pay for.

The lecture on video

▶ Play the video

In this series

• The guide: AI agents: Stanford's lecture 1, taken apart in four parts
• Part 1: An AI agent is five parts and a loop. The real question: who decides the next step?
• Part 2: BabyAGI and AutoGPT fizzled. What's left of 2023: the agent that thinks in a loop (this page)
• Part 3: AIs that debate, an AI that forgets the rule: multi-agent systems and memory, the Stanford lesson
• Part 4: MCP, the universal plug for AI agents. And a pop-up is enough to distract them

Source sheet

• Stanford Online: video "What Are AI Agents? Agentic Systems Explained | Stanford CS329Z Lecture 1 with Diyi Yang", posted on 09/11/2026, recorded on 23/09/2026.
• Stanford CS329Z: lecture 1 slides; course website and slides from the 12/10/2026 lecture.
• Shunyu Yao and co-authors: "ReAct: Synergizing Reasoning and Acting in Language Models", ICLR 2023.
• Noah Shinn and co-authors: "Reflexion: Language Agents with Verbal Reinforcement Learning", 2023.
• GitHub: AutoGPT, BabyAGI and BabyAGI Inside repositories.
• Links: the video · lecture 1 slides (PDF) · ReAct paper · Reflexion paper · AutoGPT · BabyAGI Inside

Source passages

Show the 7 sources

Ref.	Quote (original language)	URL
S1↑1↑2↑3↑4↑5↑6+17	A [00:14:13] / B [00:14:14]: "So, 2022, ReAct was proposed."; A and B [00:14:24] (A = B): "So instead of thinking once and then answering the question, what the model does here is actually a loop."; A [00:14:30] / B [00:14:30]-[00:14:33]: "It thinks about what to do next, take an action, and then look at what happened, and then use that to think again." (B: "take action"); A [00:14:56] / B [00:14:58]: "they really rely on the base model"; A and B [00:15:07]: "we have a much stronger model that could follow long instructions and plan more reliably"; A [00:15:17] / B [00:15:18]: "Within two weeks, there were actually two open-source projects came out"; A [00:15:28] / B [00:15:29]: "give the model a goal and let it write its own task list" (B: "gave"); A [00:15:40] / B [00:15:47]-[00:15:54]: "connect GPT-4 to tools such as web search, file systems, etc., and then let it keep going on on its own without a person really like typing the next prompt each time"; A and B [00:16:07] (A = B): "The reality is that none of them actually really worked very well."; A [00:16:11] / B [00:16:12]: "But they got people's attention, and really opened our imagination about what's possible."; A and B [00:16:18]: "that's also probably the time when the word "agent" really enters like everyday use"; A and B [00:19:30]: "it's really hard to make estimation into the future, because a lot of things can actually happen in two to three months these days"; A [00:19:39] / B [00:19:40]-[00:19:45]: "RSI, recursive self-improvement, and it has shown very interesting patterns on solving open math problems" (B: "showed"); A and B [00:19:52] (A = B): "I think with RSI, this pattern will get accelerated a lot, so let's see."; A and B [00:20:51]: "the API cost here is strongly associated with like the number of tokens got digested in the process" (B: "get digested"); A and B [00:21:02]: "thinking longer is actually going to lead to improved performance"; A [01:01:31] / B [00:21:30]: "The reality is that we do not know a lot about how those models are trained."; A and B [00:29:10]: "this chain of thought may not be faithful of what actually happened within LLMs" (B: "chain-of-thought"); A [00:29:40] / B [00:29:44]: "you are cooking a dish and seeing salt is out"; A and B [00:30:23]: "the dish should be savory, and since salt is out, I should find something equivalent"; A and B [00:30:40]: "the same reasoning could actually work in new, similar situations"; A and B [00:30:50] (A = B): "It also help with alignment, because you can see why the agent did what it did."; A and B [00:31:31] (A = B): "This is actually more dangerous, because you have a very convincing reasoning trace that's wrong."; A and B [00:33:53]: "Hallucination"; "ReAct is 0%" (B: "ReAct is zero"); "chain of thought is actually 56% here" (B: "56 here"); A and B [00:35:10]: "the agents are going to reflect through natural language feedback"; A [00:35:24] / B [00:35:26]: "the evaluator is going to judge: is this going to be correct or incorrect?"; A and B [00:36:31]: "compared to ReAct only, with reflection, it's definitely doing way better than the ReAct system"; A and B [00:36:41]: "it reduce hallucination a lot" (B: "reduced"); A and B [00:37:05]: "So reflection and many other similar self-reflection techniques does not involve fine-tuning or any kind of gradient updates"; A [00:37:20] / B [00:37:25]-[00:37:32]: "it actually keep using this word called "reinforce", but reinforce is actually not achieved from a reinforcement learning perspective"; A [00:37:51] / B [00:37:55]-[00:38:16]: "Sometimes you fail not because of hallucination, but because such label may have ambiguity"; "more about like interpreting what has happened so far"; A and B [00:38:54] (A = B): "it's literally just like one prompt, one pass."; A [00:38:58] / B [00:39:12]: "the model actually can revise its own errors, or getting new information to update themselves"	https://www.youtube.com/watch?v=DGI4vucR71w
S2↑1↑2↑3↑4↑5↑6	p. 25: "Oct 2022 ReAct"; "Nov 30, 2022 ChatGPT"; "Mar 14, 2023 GPT-4"; "Mar 28, 2023 BabyAGI"; "Mar 30, 2023 AutoGPT"; p. 34: "The dish should be savory, and since salt is out, I should find the soy sauce instead. It is in the cabinet to my right…"; p. 35: "Question: Aside from the Apple Remote, what other device can control the program Apple Remote was originally designed to interact with?"; "Answer: iPhone, iPad, iPod Touch"; p. 36: "Thought 3: Front Row is not found. I need to search Front Row (software)."; "Act 4: Finish [keyboard function keys]"	https://web.stanford.edu/class/cs329z/slides/lecture01.pdf
S3↑1↑2↑3↑4	p. 6, table 2: "Reasoning error Wrong reasoning trace (including failing to recover from repetitive steps) 47% 16%"; "Hallucination Hallucinated reasoning trace or facts 0% 56%"; "we randomly sampled 50 trajectories with correct and incorrect answers (judged by EM) from ReAct and CoT respectively (thus 200 examples in total)"; "there is one frequent error pattern specific to ReAct, in which the model repetitively generates the previous thoughts and actions" (text extracted from the PDF: some spaces missing)	https://arxiv.org/abs/2210.03629
S4↑1↑2↑3↑4↑5	Reflexion, abstract: "Reflexion agents verbally reflect on task feedback signals, then maintain their own reflective text in an episodic memory buffer to induce better decision-making in subsequent trials."; "Reflexion achieves a 91% pass@1 accuracy on the HumanEval coding benchmark, surpassing the previous state-of-the-art GPT-4 that achieves 80%."; table 1: "HumanEval (PY)", "80.1 (GPT-4)", "91.0"; GPT-4 technical report (OpenAI, arXiv 2303.08774), table 2: "HumanEval", "67.0%", "0-shot"; versions: "2210.03629v3", "updated 2023-03-10T01:00:17Z"; "2303.11366v4", "updated 2023-10-10T05:21:45Z"	https://arxiv.org/abs/2303.11366; https://arxiv.org/abs/2210.03629
S5↑1↑2	Releases: "autogpt-platform-beta-v0.8.1" "2026-09-24T09:35:53Z" … "autogpt-platform-beta-v0.11.2" "2026-11-11T16:51:27Z" (9 releases published after 23/09/2026, excluding the "preview-seed-fixture" pre-release); v0.9.0 (2026-10-23T01:28:07Z): "#14981 - Every block declares its effect; Auto judges cheap outward actions and asks first on costly ones"; v0.8.2 (2026-09-30T05:25:31Z): "#14209 - AutoPilot action gating modes - Ask First, Auto, Unsupervised"; first platform release: "agpt-platform-beta-v0.1.0", "2024-10-17T11:55:10Z" (verification-sources/v02-autogpt-releases-p1.json, v03-…-p2.json); repository: "stargazers_count": 188001; commits since 23/09/2026 00:00 UTC, master branch, 1 per page: "page=1213>; rel="last""	https://github.com/Significant-Gravitas/AutoGPT
S6↑1↑2	BabyAGI, README: "The original BabyAGI from March 2023 introduced task planning as a method for developing autonomous agents. This project has been archived"; last commit: "2026-01-31T05:05:45Z"; BabyAGI Inside: "created_at": "2026-11-04T21:58:08Z"; README: "BabyAGI Inside compiles a bounded BabyAGI-style agent loop into a widened Qwen3.5-2B checkpoint"; "The result is a systems demonstration—not a claim of general agency"; "All eight sequential arithmetic chains match their frozen expected results"; "Both eight-cycle structured-operation traces complete mechanically but fail exact task semantics"; "These ten traces were selected for diagnosis, not sampled to estimate accuracy."	https://github.com/yoheinakajima/babyagi; https://github.com/yoheinakajima/babyagi-inside
S7↑1↑2	Site: "4 Mon Oct 12 Agent Design Patterns & Scaffolds"; slides of 12/10, p. 35: "Now fine-tuning models to act as agents is common practice during post-training."; p. 36: "Including thoughts before actions is still very common today."	https://cs329z.stanford.edu/; https://web.stanford.edu/class/cs329z/slides/lecture06.pdf