What is an AI agent?
An AI agent is a system in which a language model decides which actions to take, executes them through tools, observes the results and repeats until a goal is met or a limit is reached. The distinguishing feature is not autonomy in the abstract but a loop: the model''s output becomes an action, and the action''s result becomes the model''s next input. Reliability falls as the number of steps in that loop rises.
The model is given a goal and a set of tool definitions — functions with typed inputs. It emits a structured tool call; the runtime executes it and returns the result; the model continues. Production systems add step limits, retries, validation of tool arguments, permission checks for irreversible actions, and persistent state so a run can be resumed or audited. Long runs also require context management, since the transcript grows with every observation.
A support agent reads a ticket, queries the order system, checks a refund policy, and either issues a refund below a threshold or escalates to a human above it — with every tool call logged.
Agents shift the engineering problem from prompt quality to systems design: permissions, idempotency, observability and failure handling. Organisations that treat agents as prompts rather than as distributed systems tend to see pilots that demo well and fail in production.
- That a single per-step success rate is reassuring. At 95% per step, a 20-step task succeeds about 36% of the time.
- That more tools improve results. Large tool sets increase selection errors.
- That agent and workflow are the same. A fixed pipeline with model calls is usually more reliable and cheaper.
- Agent benchmarks use synthetic environments and rarely capture messy enterprise systems or partial failures.
- Vendor-reported completion rates are usually measured on tasks the vendor selected.