Today, "AI agent" gets used for pretty much anything that answers a question. Yet assistant and agent describe two different roles, and the line between them has nothing to do with whether there's a conversation, or even whether an external tool gets used — it comes down to who decides the steps and drives their sequence.
We call everything an "AI agent" far too easily
A chatbot that rephrases a text isn't an agent. A tool that answers a question about some data isn't necessarily an agent either. The term has become so commonplace it now covers any moderately sophisticated conversational interface, when a fairly simple distinction actually separates several different roles.
This confusion isn't limited to marketing pitches. Media coverage alternates between "AI agents" and "agentic AI," while institutions are still working to pin down the definition. France's data protection authority, the CNIL, itself notes that no universally adopted definition exists yet. The French competition authority, the Autorité de la concurrence, nonetheless distinguishes the assistant — which responds to a request — from the agent, to which the user hands off an objective and a capacity to act — a distinction close to the one I develop here, without claiming it amounts to technical consensus.
The AI assistant: a partner for thinking things through
The AI assistant supports reflection. You hand it an idea, it rephrases it, challenges it, helps structure a line of reasoning. In its most common use, the AI assistant stays centered on the exchange: it helps you understand, formulate, decide. The user keeps the initiative and drives the conversation.
The tool-using AI assistant: the user drives the actions
Conversational assistants sometimes rely on a technical integration to look up or change external data — a file, a database, a business application. That mechanism (the MCP protocol, or other integrations presented as plugins or connectors depending on the platform) will be the subject of a separate article; here, we only keep the role it plays in this continuum.
The tool-using AI assistant carries out one or several actions directly requested or approved by the user, without driving the whole process itself. That's already more than a conversation, but each action still gets triggered by an explicit request.
The AI agent: the AI drives a process toward a goal
The AI agent isn't defined solely by its ability to use a tool. It can work out the necessary steps, call on several tools, observe their results, and adapt the rest of its action to reach a given objective. It no longer just executes a one-off request: it drives a process.
Important nuance: a one-off action — looking up a piece of data, changing another one via a tool — can very well belong to a tool-using assistant, not an agent. What defines the agent isn't the number of tools it calls, but its ability to direct and adapt a process aimed at a goal, on its own — plan, act, watch the result, adjust.
Whether it's a tool-using assistant or an agent, the level of autonomy depends on the permissions, validations, and framework set around the system. That autonomy can be tightly supervised, or, on the contrary, much broader.
Claude Code as a personal example of agentic use
Personal example: in its agentic mode, Claude Code, which I use daily to build this site, illustrates this difference concretely — it can explore the code, work out a plan of action, edit several files, run commands and tests, analyze the results, correct its approach, and possibly prepare a pull request. It's not systematic. In my own use, these actions stay bound by permissions and human oversight — but the sequence itself is driven by the AI, not replayed by hand each time.
A continuum of autonomy, not an absolute boundary
The line, then, isn't binary — it's more of a continuum across three levels. In its conversational mode, the AI assistant talks and helps you think without directly touching an external system. The tool-using AI assistant carries out one or several actions directly requested or approved by the user, without driving the whole process itself. The AI agent works out and chains the steps and tools needed on its own to reach the assigned objective, within the limits of the autonomy it's been granted.
In practice: with a tool-using assistant, you ask for each action to be carried out — you steer every step. With an AI agent, you hand over an objective, and it's the AI that works out and organizes the actions needed to reach it.
Why this distinction actually matters
This isn't just a vocabulary exercise. It shapes the question worth asking in front of any new AI tool: who decides the steps, and how far? A tool-using assistant stays under the user's direct control at every step; an agent takes on part of the decision about how to proceed. Both are legitimate — but conflating them means misjudging the actual level of delegation at play, and so the level of vigilance it calls for.
The words will keep being used loosely — "AI agent" will probably go on meaning almost anything. The distinction that matters isn't about whether there's a conversation, or a tool involved: it's about who decides the steps and drives their sequence.
Personal account, based on my day-to-day use of Claude Code. The integrations that let an assistant act on an external application — whether built on MCP or presented as plugins or connectors — will be the subject of a separate article.
Where do your own AI usages sit: assistant, tool-using assistant, agent? I'd welcome feedback or disagreement — get in touch.