AI agent
agentic system · autonomous agent
In short
An AI agent is a language model connected to tools it can call, given a goal rather than a single question, and allowed to loop — calling a tool, reading the result, deciding the next step — until it judges the goal met or hits a limit. The loop is what distinguishes it from a chatbot.
An agent is three components around a model: tools it may call, a goal stated above the level of a single action, and a loop in which it decides what to do next after each result.
Tools are ordinary functions — a database query, an HTTP request, a file write — described to the model by name, purpose and argument schema. The model never executes anything itself; it emits a structured call, your runtime runs it, and the result returns as text. An agent is therefore only as capable as the tools it is given.
The engineering that matters is almost entirely about bounding the loop:
- Step, time and token budgets, so a stuck agent stops.
- A permission boundary enforced in the runtime: read widely, write narrowly, require approval for anything irreversible.
- A replayable log of every call and result.
- Cheap verification of the outcome — the reason coding agents worked first is that tests answer "did it work" automatically.
The security consideration specific to agents is prompt injection: any content the agent reads may contain instructions aimed at it. The defence is not a better prompt but a smaller set of permissions. See the full explainer on what an AI agent is.
Frequently asked questions
- How is an agent different from a workflow?
- A workflow follows steps decided in advance. An agent decides the sequence at run time from what it observes, which buys flexibility and costs predictability.
- Why do agents fail more than expected?
- Per-step errors compound. At 95% reliability per step, a twenty-step task succeeds about a third of the time. Narrowing scope helps far more than a larger model.