Agents

A model given tools, a goal, and permission to take steps on its own.

An ordinary model answers. An agent acts: it is given a goal, a set of tools it may call — search, a database, an email client, a code repository — and a loop in which it decides what to do next, does it, looks at the result, and decides again. The word covers everything from a script that calls a model three times in sequence to a system that runs for an hour across dozens of steps.

That range is why the word is nearly useless in a vendor conversation. The useful questions are narrow: what tools can it reach, what can it change rather than only read, how many steps can it take before a person sees the result, and what happens when a step fails halfway through.

Agents fail differently from chatbots. A chatbot that gets something wrong produces a bad paragraph. An agent that gets something wrong produces a bad action, and may then take nine more actions on top of it.

Why it matters here

Delegation is the whole governance question. In a public sector setting an agent that can read is a search problem; an agent that can write, send, pay or decide is an accountability problem, because someone must still be answerable for the act. The control that matters is not the model — it is the permission boundary around the tools, and whether the audit trail reconstructs what happened.

The question to ask

Which actions can this take without a person approving them, and can we reconstruct every step afterwards?

Reviewed 2026-09-20 · All decoders