An AI agent, in the definition Simon Willison settles on, is a large language model that runs tools in a loop to achieve a goal — the widely-used “tools in a loop” formulation, extended by a goal clause he adds himself. The model is given the ability to request actions from its harness; the outcome of each action is fed back into the model, which continues reasoning and acting until the goal is reached. Each part does work: “tools in a loop” names the mechanism, and “to achieve a goal” supplies the stopping condition that distinguishes it from an unbounded loop.
The formulation has two recorded moments in these sources. In May 2025, at an Anthropic developer conference, an Anthropic speaker is reported as stating that at Anthropic “agents are models using tools in a loop” — an account written up approvingly by Simon Willison, who had been frustrated that the word was used constantly at the event without being defined. In September 2025 Willison adopted that formulation as his own working definition, adding the goal clause to it himself after considering and rejecting a narrower “a goal set by a user”.
Usage
The longer treatment sets out what the definition includes and deliberately leaves out. Memory is not a separate requirement: the source argues the tools-in-a-loop model has a basic form of memory built in, since the tool calls accumulate in a conversation with the model and the earlier steps supply the short-term memory needed for the current goal, with long-term memory best added as a further set of tools. Nor does the goal have to come from a user — the source considers and rejects “a goal set by a user” as a necessary clause, on the grounds that sub-agent patterns already exist in which one LLM sets another’s goal. This is the shape described under Sub-agent architecture.
Restating the definition for a general readership in Agentic Engineering Patterns, Willison spells out what the word “agent” actually denotes in it, which the compressed formulation leaves implicit: the agent is the software that calls an LLM with your prompt and a set of tool definitions, then executes whatever tools the LLM requests and feeds the results back into it. On that reading the model is a component of the agent rather than the agent itself — the same separation the harness framing makes (see AI Coding Agent). He adds the specialisation that produces a coding agent: its tools include one that can execute code. He also places the difficulty of defining the term further back than the LLM era, noting that clearly defining “agent” has frustrated AI researchers since at least the 1990s.
The source is equally clear about what it rejects. It singles out the definition of agents as replacements for human staff as its least favourite, and argues that category “remains science fiction” — the distinguishing feature of human staff being accountability, the capacity to take responsibility and learn from mistakes, which it illustrates with a quoted maxim that a computer can never be held accountable and therefore must never make a management decision. It notes the related point that humans have agency in the sense of forming their own goals, which AI agents, despite the name, do not. It also names OpenAI as the largest single source of confusion, observing that the company’s CEO describes agents as systems that do work independently, that its “ChatGPT agent” feature is a browser automation system, and that only its Agents SDK closely matches the tools-in-a-loop idea.
The vendor’s own distinction
Building effective agents is Anthropic’s earlier and longer treatment of the same territory, published in December 2024, before the May 2025 remark above. It approaches the definitional problem by not settling it. Anthropic notes that customers define “agent” in several ways — some meaning fully autonomous systems operating independently over extended periods, others meaning prescriptive implementations following predefined workflows — and rather than choosing, it groups all the variations under agentic systems and draws an architectural line inside that category. Workflows are systems where LLMs and tools are orchestrated through predefined code paths; agents are systems where LLMs dynamically direct their own processes and tool usage, keeping control over how they accomplish a task.
The line is therefore drawn at who decides the path, not at how capable or autonomous the system is, and it is what lets the post attach different advice to each side: workflows for predictability and consistency on well-defined tasks, agents where flexibility and model-driven decision-making are needed at scale.
The post’s description of an agent in operation matches the loop formulation closely. An agent begins from a human command or an interactive discussion, then plans and operates independently, possibly returning to the human for information or judgement; during execution it is described as crucial that the agent gain “ground truth” from the environment at each step — tool call results, code execution — to assess its progress. Anthropic states the implementation plainly: agents are “typically just LLMs using tools based on environmental feedback in a loop”, which is the same shape as the definition above, with the environment supplying what keeps the loop honest.
Where it adds something the loop formulation leaves open is termination. A task often ends on completion, but the post describes it as also common to include stopping conditions such as a maximum number of iterations to maintain control — a bound from outside, alongside the goal that bounds it from within.
A research-literature checklist
From LLMs to LLM-based Agents for Software Engineering: A Survey of Current, Challenges and Future approaches the same definitional problem from the research side, and reports that no comprehensive and widely accepted definition yet specifies the exact capabilities an LLM must exhibit to be considered an LLM-based agent, particularly within software engineering. Rather than settle on a sentence, that survey synthesises criteria from mainstream definitions and from first-half-2024 literature and offers six of them, treating an architecture as an agent when it meets most:
- The LLM serves as the brain — the centre of information processing and the generation of thought.
- The framework possesses decision-making and planning abilities, not only the language understanding and generation capabilities of the LLM.
- Where tools are available, the model can autonomously decide when and which tools to use and integrate the results into its predictions.
- The model can select the optimal solution from among several homogeneous results.
- The model can handle multiple interactions and maintain contextual understanding.
- The model has autonomous learning capabilities and adaptability.
The survey treats the first four as fundamental to a basic level of agency and the last two as more advanced behaviours that are beneficial but not strictly necessary for an initial classification, which it presents as giving flexibility while preserving rigour. It then applies the list as a working test: a self-improvement framework whose decision-making and planning are confined to code generation and improvement, with no overall planning and a completely fixed execution sequence, is classified as an advanced LLM application rather than an agent, while a multi-role collaboration framework is classified as an agent on the strength of its role distribution, self-improvement ability and autonomous decision-making even though it lacks external tool integration. The two judgements show where the survey puts the line: on whether the control flow is the model’s to decide, not on how elaborate the surrounding scaffolding is.
When It Applies
The definition is offered as a working convention among technical implementers, not as a settled industry meaning, and the source is explicit about the boundary: when a technical implementer says “agent” it proposes reading them as meaning tools wired to an LLM to achieve goals in a bounded loop, while outside that field — in conversation with non-technical business audiences — it advises clarifying which definition the other person is using before proceeding.
Its stated rationale is that jargon is only useful when both parties share a definition, and that a contested term makes communication less effective rather than more. The source treats this as a long-standing problem rather than a new one, citing an earlier remark by Carl Hewitt that the question “what is an agent?” embarrasses the agent-based computing community much as “what is intelligence?” embarrasses mainstream AI. How well-established the definition is can be gauged from the source’s own account of arriving at it: the author reports having previously collected and attempted to cluster 211 crowdsourced definitions, and presents this one as a definition he became willing to use without scare quotes rather than as an authority’s ruling. Compare Semantic Diffusion.
Related Terms
AI Coding Agent, Augmented LLM, Sub-agent architecture, Agentic Coding, Long-running Agent, Semantic Diffusion, I think “agent” may finally have a widely enough agreed upon definition to be useful jargon now, From LLMs to LLM-based Agents for Software Engineering: A Survey of Current, Challenges and Future