Build for measured behavior.

Separate durable instructions from task input. Use strict schemas for machine-readable output, narrow tool permissions, explicit confirmation for consequential actions, and representative evaluations before changing models or prompts.

Production checklist

1. Keep secrets server-side and tightly scoped.
2. Evaluate model or prompt changes before rollout.
3. Budget input, output, retries, and tool calls.
4. Log redacted request IDs, latency, and failures.
5. Require citations or abstention for retrieved evidence.

Official documentation ↗

← All guides