0/6 checked
Understand
Before anyone relies on the agent, prove it gives the right answers to the questions real people will ask. Test as you build with the debugger or Playground, then measure formally with evaluation sets, and get go-live approved.
Key points
- Workflows have a . Supervisor agents have the .
- An holds your questions, the answers you expect and pass marks.
- Oracle says to re-run evaluations after every change, and after a model update.
- Record the approver's role and the exact evaluation run they reviewed.
Skip ahead (optional):
Leila Fielding