AI Agent
An AI agent is a model given tools and the right to act in a loop: it decides what step to take, takes it, looks at the result and carries on until the task is done.
How it works
An ordinary model answers once to one request. An agent is given a goal and a set of tools — search, file access, API calls, running code — and works in a loop: choose an action, take it, judge the result, choose the next.
The key difference is that the plan is not fixed in advance. The model decides how many steps are needed and which, so the same task can be solved differently from run to run.
Hence the central difficulty: an early mistake propagates through the whole loop. An agent that misread the task will confidently and diligently do the wrong thing, and what stops it has to be an external limit rather than its own judgement.
What matters in practice
- Limit the permissionsRead access and write access are different things. Irreversible actions should require confirmation.
- Cap the number of stepsWithout a limit an agent can settle into a long loop, spending time and money on useless work.
- Check the intermediate stepsA final answer can look convincing with an error in the middle. Look at what it actually did.
- A narrow task beats a broad oneAn agent with one clear goal and three tools errs far less often than a general-purpose one.