English pilot — the directory is mainly in French for now.
No need to be a developer to understand the mechanics: every agent, from the simplest to the most sophisticated, runs in the same loop. Here it is, step by step.
Goal → 1 · Perceive → 2 · Think → 3 · Act → 4 · Verify ↺ → Result
The agent keeps looping until it reaches the goal, runs out of its maximum number of steps — or asks you for help.
1 · Perceive. The agent receives a trigger (an email arrives, a time of day comes, you submit a request) and gathers the context: the message, your files, the history of the task.
2 · Think. The language model breaks the goal down into steps, chooses the right tool and, if needed, drafts the content (a reply, a query, a plan).
3 · Act. The agent calls a tool: send the email, query a database, click in a browser, create a file. This is the moment the AI stops talking and starts doing.
4 · Verify. The agent looks at the result of its action: success, error, new information. It adjusts the next step — or starts over differently. This verification is what distinguishes an agent from a simple script: it reacts to what actually happens.
At the center, a large language model (GPT, Claude, Gemini, Mistral…). It understands requests, reasons and writes — but on its own, it can do nothing but produce text. Everything beyond conversation goes through the tools.
A tool is a function the agent can call: “read this email”, “create this event”, “search this database”. To keep every piece of software from inventing its own plug, an open standard has taken hold: MCP (Model Context Protocol), which lets you connect thousands of existing tools to almost any agent. Our profile MCP (Model Context Protocol) explains how to use it, and the glossary defines the neighboring terms (API, A2A…).
In the short term, the agent keeps the thread in its “context window”: the task in progress, the latest exchanges. In the long term, some agents keep notes, preferences or reference documents that they reread on every assignment — which gives the impression that they “learn” your habits. In practice, they reread very good notes.
The loop can go off the rails: an agent that verifies poorly can loop, repeat an expensive action or pursue a badly worded goal with great conviction. That's why serious systems add limits inside the loop itself: a maximum number of steps, capped budgets, human approvals at sensitive points. That's the subject of the guide AI agents: risks and guardrails — and, on the building side, of our profile LiteLLM.
Browse the directory's 111 agent profiles →