AI Vocabulary

Assistant or agent: which one actually does the work?

An assistant answers a request and stops. An agent pursues an objective: it breaks the task down, chains several operations, consults sources, and delivers a result without being told each step. The difference is not one of power but of autonomy.

What an assistant does

You give it an instruction, it produces an answer. Rephrasing a paragraph, explaining a concept, summarising a text you supplied. Each exchange is closed: what you get depends solely on what you wrote.

The relationship is one of question and answer. If the task takes five steps, you chain them, in five successive requests, carrying the result of each into the next.

What an agent does

You give it an objective, it works out the steps itself. On a request such as check whether this contract is consistent with our usual position on liability, an agent opens the contract, identifies the clause, looks for earlier positions, compares them, and returns a discrepancy.

Three things distinguish it: it divides the task without being told to, it uses external tools — a search, a database, a file — and it continues until the objective is met or it finds that it cannot be.

Where the boundary actually falls

The useful criterion is not the product name but this question: who decides on the second step? If it is you, by writing a fresh request, you have an assistant. If it is the system, from the result of the first, you have an agent.

A second criterion helps: does the agent use external tools on its own initiative? Opening a file, running a search, consulting a database are actions; a system that triggers them without being asked each time acts, one that waits for instruction answers.

This explains a common confusion. Many products presented as agents are assistants with shortcuts added: the division was written in advance by the designer, it does not depend on what the system observes along the way. That is automation, which is useful and different.

What it does not solve

Autonomy increases the reach of an error rather than reducing it. An assistant that errs produces a false answer you read; an agent that errs at the second step builds the next three on that error, and you see only the final result.

Autonomy is also paid for in visibility. The more steps the system chains alone, the less you see what it did, and the harder final verification becomes: you receive a conclusion without the path to it.

An agent knows no better what it does not know. If it fails to find the earlier position it was looking for, it may conclude that none exists and carry on, when it was simply filed elsewhere.

And autonomy does not replace context. A highly capable agent that does not know your matter will make division decisions consistent with what it believes the request to mean, not with what the matter requires.

Why it matters to a lawyer

Because supervision arises differently in each case. Reviewing an assistant's answer means reviewing a text. Reviewing an agent's result means checking a chain whose steps you did not see.

The requirement follows: an agent must be able to show its path. Not to explain its reasoning, which no system does reliably, but to state what it opened, what it retained, and at what point it stopped looking.