What is an AI agent? A model, tools, a loop and a human at the gate
- glossary
- ai-agents
An AI agent is a language model with tools and a loop: it reads a task, decides the next step, acts, checks the result, and goes round again until the job is done or a person has to answer. Seventh glossary entry, and the one where the marketing is loudest. Most "agents" I get shown are a chat window with a new label.
- The plain definition. A model that picks the next action, tools that carry it out, a loop that feeds the result back, and a rule for when to stop.
- The one-minute test. Ask what it can touch without you. Nothing: a chat. Files, a browser, a database: an agent.
- Autonomy is a dial. The same calls every morning is a workflow. Picking the next call from the last result is an agent.
- Where the risk lives. In the tools, not the model. Same model, read-only key versus database password: two different systems.
- What the label hides. The stop condition and the human gates. A demo shows the loop running. Ask to see it stop.
Most of my agents run inside Claude Code: I type a job, it reads and edits files, runs commands, checks the output, and comes back when the work passes or when a decision is mine to make.
The plain definition
An AI agent is a large language model given tools and a loop. The model reads the task and the current state and picks the next action, a tool runs it through function calling, and the result goes back into the model's context. That repeats until a stop condition fires: the work passes its check, the budget runs out, or the next action needs a person. People sit at gates: before a write, a spend, or a message that leaves the system.
Every part is something you can point at: a model name in a config, a tool list, loop code, a stop rule, hooks at the gates.
Google Cloud calls agents "software systems that use AI to pursue goals and complete tasks on behalf of users". IBM leads with autonomy: "autonomously performing tasks on behalf of a user or another system". The one I use is LangChain's, because it is testable: a system that "uses a large language model to decide the control flow of an application". If the model decides what runs next, it is an agent. If code decides and the model fills in text, it is not.
What is an AI agent in layman's terms?
A chatbot answers. An agent does. Give a chatbot a question and you get text back. Give an agent a job and it goes and looks, tries something, sees what happened, and tries again until the job is done or it has to ask you.
What it means in practice
Three parts carry the weight day to day: the loop, the tools and the stop rule.
What is an AI agent and how does it work?
An agent works as a loop. Task in, plain language. The model plans the steps. It acts through tools: it edits a file, runs a command, calls an API. Then it verifies: tests, output, logs. Not done yet, go again. Done, stop and report.
task -> plan -> act -> verify -> not done yet? go again
done, or blocked? stop and report
That loop is the orchestration layer, the second of the six layers I draw for an agent architecture. The other layers exist because a loop left alone will run forever, spend anything, and trust everything it reads.
The tools are where an agent stops being a text generator: read a file, query a database, open a page, send a message. Mine mostly arrive through MCP servers, a standard way to hand a model callable tools.
A stop condition is what separates an agent from an incident. Anthropic's guide to building agents notes it is common to add "stopping conditions (such as a maximum number of iterations)". Mine are written before the run: enough evidence, data unavailable, or a cap on tool calls. Each of the seven agent examples I run lists its stop condition next to its human gate.
Autonomy is a dial. Anthropic separates workflows, "orchestrated through predefined code paths", from agents, where "LLMs dynamically direct their own processes and tool usage". LangChain adds that "Autonomy is a spectrum". Turn the dial up only when a fixed workflow is harder to test than the agent.
What people get wrong
| People say | What is true |
|---|---|
| The vendor calls it an agent | The label is free. The tool list and the stop rule are the spec |
| The model is the agent | The model is one box. The agent is the tools, loop, stop rule and gates around it |
| A smarter model means a safer agent | Safety comes from what the agent can touch, not from how well it reasons |
| Autonomous means unattended | My useful agents stop at a gate before anything is written, spent or sent |
Is ChatGPT an AI agent?
The chat window is not. You type, it answers, nothing outside the conversation changes. That is an interface to a model. OpenAI's July 2025 announcement page for ChatGPT agent describes a mode in which the product runs tasks "using its own virtual computer". A mode like that meets the definition: model, tools, loop.
The label belongs to the mode, not the brand.
What are the 5 types of AI agents?
The five come from the standard AI textbook by Russell and Norvig, which lists four basic kinds of agent program plus learning agents as a fifth: simple reflex, model-based reflex, goal-based, utility-based, and learning agents. The list was written for agents in general, not language models. It is a fine exam answer and not how anyone I know runs one.
I sort agents by what they can touch instead: read-only, writing inside a sandbox, or reaching production and other people. The textbook has no column for blast radius.
One example from my own work
The pipeline that publishes this site is the cleanest example I have. Every post starts as a file with a draft flag set to true. An agent researches, writes the draft, and checks its links, claims and metadata against the contract. A second model with different instructions tries to refute it, and that judge has caught real errors before publication. Then I read it and flip one line in the file.
The deploy is a separate run. From one instruction, the agent builds the static export, drops the pages not ready to publish, pushes the files to the server, and checks that the live URLs return HTTP 200. That publish step does not write and does not choose. It reveals what a person already approved. The build refuses to run if a post's front matter drifts from the contract.
Agent proposes, judge checks, human approves, publish step executes. The human is gate six and the judge is gate seven of the twelve I run before an agent gets real access. The loop, the tool list, the stop rule and the flag are the agent.
Where to go deeper
Entry one, What is AI, drew the big circle. Entry two put the engine inside it. Next in the glossary: what happens when the model needs facts it was never trained on.
Sources: LangChain, What is an AI agent? (Jess Ou, July 31, 2026); Anthropic, Building effective agents (December 19, 2024), both read on 30 August 2026. Google Cloud, IBM and OpenAI are quoted from their own pages via search results that day; those did not load in full, so no links. The agent types come from Russell and Norvig, Artificial Intelligence: A Modern Approach, chapter 2, cited from reference knowledge rather than a source read that day.