Build an agent

The loop #

1. Send the user task and your allowed tool schemas.
2. If the model returns text, present the result.
3. If it requests a tool, validate the name and arguments.
4. Check permissions and obtain approval for external effects.
5. Execute the permitted action and append the tool result.
6. Continue until completion or a time, step or spend limit.

Keep reliable state #

Record tool call identifiers and confirmed effects outside the model context. Avoid executing the same non-idempotent action twice after a timeout. Distinguish an incomplete inference response from a completed tool operation.

Use the correct protocol #

Start with Tool calling and Streaming. Keep Messages, Chat Completions and Responses tool formats separate.

Observe cost and failures #

Keep x-zrelay-request-id, bound retries, and inspect request history when the outcome or charge is unclear.