# Harness evaluation checklist

- [ ] Define the outcome and how it will be checked.
- [ ] List the information sources and allowed tool actions.
- [ ] Decide what memory should persist and what is task-specific.
- [ ] Set approval boundaries and stopping conditions.
- [ ] Try incomplete input or a blocked tool and inspect the failure.
- [ ] Open the result and compare it with the original request.

Record the task, product version, model, permissions, sample data and results. Compare alternatives with the same task and data.
