Agents

A request-based Agent handles a command and returns a response. The interaction can include tool calls before the model produces its final answer. Use it for focused request-and-response behavior or as a step in a workflow whose sequence is defined in code.

An Autonomous Agent works on typed tasks in a durable decision loop. The model selects its next action until it produces a result, reports failure, or reaches an iteration limit. Tasks have identities and results that clients can inspect independently of the running agent.

Choose between them based on who decides the sequence: your code or the model. Autonomous Agents also provide coordination capabilities for delegation, handoff, teams, and moderation. They already persist task progress; a workflow is useful when it owns additional process steps, not merely as a durability wrapper.

A model sees the prompt and context supplied for each call. Session memory supplies interaction history. Tools retrieve domain data or perform actions; the runtime executes them after any applicable tool-call checks.

Structured output constrains the response shape, but does not establish that its content is correct. Guardrails control permitted actions and replies. Evaluators assess behavior against quality criteria, before release and on live traffic.

Model calls add latency and token cost, and similar inputs can produce different results. Bound tool use and iterations, keep context focused, and evaluate the complete interaction rather than one exact response. More autonomy also means more possible execution paths to test.