Reference · How-to · ~7 min
How to ship an agent with an eval gate
Last updated
No agent reaches prod without a golden task set.
No agent reaches prod without a golden task set.
Steps
1. Collect 10 real tasks your agent must pass
2. Cap loop at 5 steps; block dangerous tools
3. Run eval set in CI on every prompt/tool change
4. Log traces for failures (see tracing how-to)
5. Canary 5% traffic; watch tool error rate