01
The multi-system test
An agent adds value when a task requires gathering information from several places, deciding what to do, and acting — and when the sequence varies by case. If the sequence is fixed, write the pipeline.
02
Permissions are the product
The design work in an agent project is mostly about boundaries: which tools it may call, what it may write, when it must ask, and what gets logged.
- Read-heavy tools open, write tools gated
- Approval checkpoints for irreversible actions
- Every tool call logged with inputs and outputs
03
Evaluate on real cases
Build a set of historical cases with known good outcomes and re-run them on every change. Without that harness you cannot tell whether a prompt change was an improvement or a regression.