An AI agent is five parts and a loop. The real question: who decides the next step?

In short

• For Diyi Yang, an agent comes down to five parts: a large language model at the centre, planning and reasoning, memory, tools, and the environment it acts in.
• The idea isn't new: she traces the first use of the word in AI back to a 1973 paper about programs that act on their own.
• Workflow: the code fixes the steps. Agent: the model picks the next one.

1973: programs that wait for a message

The word agent comes from Latin: one who acts. In AI, Diyi Yang places its first indexed use in a paper by Carl Hewitt, Peter Bishop and Richard Steiger, presented in 1973 at the IJCAI conference. I opened the original PDF: an actor is defined there as "an active agent which plays a role on cue according to a script". Software that acts on its own, more than fifty years ago.

Then it all rolls by: Marvin Minsky's "Society of Mind" (1986), where small sub-agents work together; the textbook definition from Russell and Norvig (1995), for whom an agent is anything that perceives through sensors and acts through actuators; reinforcement learning, with AlphaGo beating humans at Go. Agents, yes, but each one trained for a single task.

What changes everything, in her view, is large language models: they can reason step by step and call tools for what they can't do alone. Those are the two prerequisites for agents. What came next, ReAct in 2022 and then the 2023 rush with BabyAGI and AutoGPT, is part 2.

The pepper shaker, taken apart

The example comes from ReAct, a 2022 method. The model itself is the core. When it writes "I need to find a pepper shaker, more likely in cabinets", that's planning and reasoning. Everything it has done so far is its memory. Going to the first cabinet, taking the pepper shaker: tool calls. And what it sees after each move is the environment answering back. "You can see that at least there are five components if we just look at this very simple example from ReAct," she sums up.

Part by part, according to the lecture:

• The core: a stronger model generally makes a better agent. Hence a debate, she says: train the model for agent work, or improve the harness, the code around it.
• Planning: breaking a big task into small pieces, sometimes handed to sub-agents.
• Memory: so that an agent working for more than two hours still knows what has happened.
• Tools: a model that's strong at maths doesn't know which room the class is in today; with a search on Stanford's website, it does. And there was a time when GPT-4 failed at multiplying two three-digit numbers.
• The environment: what the agent observes and where it acts: a computer, a web browser, apps, games.

"So what you see on the screen right now, this is actually a skeleton of an agent," she says, one drawing among others. Other researchers use other words, but for her they all describe the same thing: a model that reasons, remembers, uses tools and acts in a loop within its environment.

Workflow or agent: who holds the wheel?

Chapter "Compound AI systems: workflows vs. agents" (53:31). Real products combine a language model with other parts: compound AI systems. Diyi Yang asks the question that sorts everything: who decides the steps?

In a workflow, the code does. Her example: Agentless, which fixes a bug reported on GitHub in fixed stages: find the file, write the patch, test. The model works inside each stage but never picks the next one: it's a script.

In an agent, the model decides for itself. SWE-agent chooses which files to open, what to run, and when it's done. More flexible, and more responsibility for the model.

A recipe followed to the letter, versus a cook who opens the fridge and improvises. (Kira asked me which one she'd want at the stove. Depends on the dinner.) And the workflow is nobody's poor relation: AlphaCode 2 generates up to a million programs for one coding problem, then filters and scores them.

The most powerful isn't the most reliable

Chapter "Key challenges: reliability, safety & privacy" (1:01:05). Diyi Yang shows a Princeton project on agent reliability: same result if you rerun the same task? Resistance to small variations? Failures you can understand and recover from? Ranked by accuracy, the models line up; their reliability scores follow no trend. The most powerful agent may not be the most reliable one.

I opened the Princeton dashboard cited on the slide (12/11/2026, 11:41 UTC). Across two combined tests, GAIA (questions that require searching the web) and τ-bench (an airline customer service), Gemini 3.5 Flash has the best accuracy, 80.0%, and a reliability of 0.85 out of 1. Claude Opus 4.5 solves fewer tasks, 74.6%, but it's the most reliable in the ranking, 0.87. This score out of 1 is the average of three marks: consistency from one run to the next, predictability, robustness; safety is scored separately.

My take: the more you hand the wheel to the model, the more you need to check that it does the same thing twice. Accuracy alone won't tell you.

Words to remember

• AI agent: an AI that observes, decides, acts and starts again, instead of just answering.
• Large language model: the program that produces text, at the heart of the agent (GPT, Claude, Gemini…).
• Workflow: steps fixed by code; the model works inside them without choosing what comes next.
• Harness: the code around the model that turns it into an agent.

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? (this page)
• Part 2: BabyAGI and AutoGPT fizzled. What's left of 2023: the agent that thinks in a loop
• 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.
• Carl Hewitt, Peter Bishop and Richard Steiger: "A Universal Modular ACTOR Formalism for Artificial Intelligence", IJCAI 1973.
• Princeton: AI agent reliability dashboard and the paper "Towards a Science of AI Agent Reliability".
• GitHub: SWE-agent and Agentless repositories.
• Links: the video · course website · lecture 1 slides (PDF) · 1973 paper (PDF) · Princeton dashboard

Source passages

Show the 6 sources

Ref.	Quote (original language)	URL
S1↑1↑2↑3↑4↑5↑6+13	A and B [00:08:35]: "So in AI, the first indexed use of it is in this specific IJCAI paper from 1973,"; A [00:08:35] / B [00:08:44]: "Hewitt and Bishop and Steiger"; A [00:09:18] / B [00:09:24]: "how a group of sub-agents actually work together to get something done"; A [00:10:03] / B [00:10:10]: "anything that perceive through sensors and act through actuators" (B: "acts"); A and B [00:11:12] (B: "OpenAI 5"): "AlphaGo and OpenAI Five"; A [00:11:12] / B [00:11:18]: "this kind of approach actually can reach superhuman performance"; A [00:11:27] / B [00:11:26]: "each of them are actually trained for a specific task"; A and B [00:12:28]: "they can also reason step-by-step"; A [00:12:28] / B [00:12:33]: "use tools for the things that they cannot do alone"; A and B [00:14:02]: "Those are two very, very important prerequisite for the emergence of agents."; A [00:22:14] / B [00:22:15]: "this particular task is to put a pepper shaker in a drawer"; A and B [00:22:31] (A = B): "I need to find a pepper shaker, more likely in cabinets."; A and B [00:23:09] (A = B): "You can see that at least there are five components if we just look at this very simple example from ReAct."; A [00:23:38] / B [00:23:42]: "So stronger model in general means better agents."; A and B [00:23:46]: "should I think about training for agents or should I just care about better harness or other improvement?"; A and B [00:24:03]: "they actually need to decompose them into smaller pieces and then maybe have sub-agents to deal with each of them"; A and B [00:24:19]: "as agents work on the task for more than two hours"; A [00:24:34] / B [00:24:39]: "where is CS 329Z happening today on campus, probably models cannot answer that"; A [00:25:23] / B [00:25:22]: "it actually cannot solve 3 digits times 3 digits" (B: "three digits times three digits"); A and B [00:26:08] (A = B): "So what you see on the screen right now, this is actually a skeleton of an agent."; A [00:26:13] / B [00:26:14]: "this is only one way to draw it"; A and B [00:27:16]: "The left side is from an OpenAI researcher back in 2023."; A and B [00:27:42]: "it really kind of point to the same architecture here"; A [01:06:00] / B [01:01:00]: "Key design choice at the system layer is: who decides the steps?" (B: "A key design choice at the system layer is who decide the steps"); A [01:06:19] / B [01:01:19]: "Both are what people refer to as compound AI systems, that is a large language model plus other components."; A [01:06:39] / B [01:01:39]: "Examples such as Agentless, which fix a GitHub issue in a very fixed set of stage, like finding the file, writing patch, and doing test"; A [01:06:51] / B [01:01:51]: "So the large language model works inside each stage, but they never pick the next one."; A [01:07:01] / B [01:02:02]: "In the agent setting, the large language model decide for itself."; A [01:07:06] / B [01:02:07]: "in SWE-agent, it actually choose which files to open, what to run, and when it's done"; A [01:07:30] / B [01:02:30]: "workflow system can actually be very, very strong"; A [01:02:21] / B [01:07:22]: "consistency: if you let the agent run multiple times, do they give you the same outcome?"; A [01:03:00] / B [01:08:00]: "if you look at the reliability, those scores do not have a trend"; A [01:03:12] / B [01:08:12]: "The most powerful agent may not be the most reliable agent"	https://www.youtube.com/watch?v=DGI4vucR71w
S2↑1↑2↑3↑4↑5↑6+7	p. 19: "First indexed “AI use” is from Hewitt, Bishop and Steiger in 1973"; p. 20: "Marvin Minsky on the Society of Mind in 1986"; p. 21: "Russell and Norvig's textbook (1995) defines an agent as anything that perceives through sensors and acts through actuators"; p. 29: "Anatomy of an Agent (ReAct: Yao et al., 2022)"; "LLM Core"; "Planning & Reasoning"; "Memory"; "Tools"; "Environment"; p. 30: "Desktop OS, Web browser, Mobile apps, Games, Databases"; p. 31: "lilianweng.github.io/posts/2023-06-23-agent/"; "Cognitive Language Agents (Sumers and Yao, 2024)"; p. 56: "Workflow: code fixes the steps (e.g., Agentless, AlphaCode 2)"; "Agent: the LLM decides the steps (e.g., SWE-agent)"; "It generates up to 1 million solutions for a coding problem then filters and scores them."; p. 62: "Consistency across runs"; "Robustness to perturbations"; "Predictable failure modes"; "hal.cs.princeton.edu"	https://web.stanford.edu/class/cs329z/slides/lecture01.pdf
S3↑1↑2↑3	p. 1: "A Universal Modular ACTOR Formalism for Artificial Intelligence"; "Carl Hewitt Peter Bishop Richard Steiger"; "Intuitively, an ACTOR is an active agent which plays a role on cue according to a script" (read on the image of the scanned page, p. 235 of the proceedings: "plays a role on cue according to a script."; the PDF's OCR gives "role-Dn cue", "script~")	https://www.ijcai.org/Proceedings/73/Papers/027B.pdf
S4↑1↑2	Site: "1 Wed Sep 23 Foundations & Landscape Introduction — What Are Agentic Systems?"; "4 Mon Oct 12 Agent Design Patterns & Scaffolds Workflows vs. agents and the design patterns for each"; slides of 12/10, p. 10: "Workflows are more predictable than agents because the control flow lies in the program logic."; "Agents are more adaptable/flexible than workflows because the control flow lies in the LLMs decisions."; p. 12: "Real systems often combine workflows and agents."	https://cs329z.stanford.edu/; https://web.stanford.edu/class/cs329z/slides/lecture06.pdf
S5↑1↑2↑3↑4	Dashboard, "All Benchmarks" list: "1 Claude Opus 4.5 74.6% 0.87"; "3 Gemini 3.5 Flash 80.0% 0.85"; "We evaluate 15 agents across 2 benchmarks on twelve metrics spanning four reliability dimensions"; "GAIA Real-world question-answering tasks requiring multi-step reasoning, tool use, and web browsing."; "Curated 26-task subset of the airline domain"; methodology: "The overall reliability score is the arithmetic mean of three dimension scores"; "Safety is reported separately and not included in the overall aggregate"; arXiv: "2602.16666v3", "updated 2026-06-02T02:15:05Z", "Towards a Science of AI Agent Reliability"	https://hal.cs.princeton.edu/reliability/; https://hal.cs.princeton.edu/reliability/methodology/; https://arxiv.org/abs/2602.16666
S6↑1↑2	SWE-agent: "tag_name": "v1.1.0", "published_at": "2025-05-22T16:11:39Z"; commits since 23/09/2026: "[]"; Agentless: last commit "2024-12-22T19:29:28Z"; SWE-agent README (read on 12/11/2026, 12:01 UTC): "Most of our current development effort is on mini-swe-agent, which has superseded SWE-agent."; mini-SWE-agent: "tag_name": "v2.4.6", "published_at": "2026-07-23T03:12:33Z"; commits since 23/09/2026: "[]"	https://github.com/SWE-agent/SWE-agent; https://github.com/OpenAutoCoder/Agentless