English pilot — the directory is mainly in French for now.
Twenty-two terms, one short, clear definition each — the ones that come up in every profile in this directory and in every article about agents.
Software you hand a goal to, which chains the steps to reach it on its own, using tools. See the guide What is an AI agent?
An agent able to work for a long time without supervision between two human interventions. The greater the autonomy, the stricter the guardrails must be.
An open protocol (Google, 2025; version 1.0 in March 2026) that lets agents from different providers discover each other and delegate tasks to one another, via “Agent Cards” describing their capabilities. Where MCP connects an agent to tools, A2A connects agents to each other. See the profile MCP (Model Context Protocol).
An interface that lets one piece of software use another in a structured way (read emails, create an event…). APIs are how agents act on your tools.
A technical password that identifies your account with a model or service provider. Whoever holds it can spend your credit: protect it and cap it.
An agent's ability to use a computer like a human: see the screen, click, type. Useful when a piece of software has no API, but slower and more fragile.
A technique where several models answer the same question, review each other anonymously and have a verdict synthesized — for important decisions. See the profile LLM Council.
The amount of information a model can keep “in mind” during a task: the conversation, the documents provided, the results of the steps. It is the agent's working memory.
The event that launches the agent: an email arriving, a time of day, a form being filled, a sentence you type.
A limit set on an agent: maximum budget, number of steps, human approval, restricted access. See the guide AI agents: risks and guardrails.
A false answer produced with confidence by a model (an invented fact, an imaginary source, an approximate figure presented as exact). The countermeasure: require sources and verify.
The “brain” of today's agents: a model trained on immense amounts of text (GPT, Claude, Gemini, Mistral…) that understands and produces language.
An open standard that lets you connect tools and data to almost any agent, via “MCP servers”. It has become the universal plug for agents. See the profile MCP (Model Context Protocol).
What the agent keeps between tasks: notes, preferences, reference documents. Not to be confused with the context window, which is its immediate working memory.
The AI program that reasons and writes. A single agent product can often run on several models, at different prices and levels.
The art of organizing the work of several agents or steps: who does what, in which order, with which controls. Frameworks such as CrewAI or LangGraph exist for this.
A function an agent can call to act: send an email, search the web, edit a file. Without tools, a model only chats.
The instruction given to the agent or the model: the request, its context and its constraints. The quality of the prompt directly influences the quality of the result.
A technique that lets an agent look up its answers in YOUR documents (rather than in its training memory alone) and cite them. It's what makes answers verifiable.
An isolated environment where the agent works without being able to touch the rest of your system. Essential for agents that handle files or code.
A small chunk of text (roughly a word or less) used as the unit of account by models: API calls are billed by the number of tokens read and written.
A sequence of steps defined in advance. A workflow is predictable; an agent can follow a workflow, but can also deviate from it intelligently when the situation calls for it.
Browse the directory's 111 agent profiles →