Set the bar before you test
Write these into the agent specification before the first test runs, so the results are judged against a bar nobody moved:- The evaluation threshold the agent has to meet.
- Autonomy for each action: what it may do on its own, what it drafts for a person, and what it never does.
- Red lines: actions it must refuse.
What to test
Run the tests against approved test data and keep the evidence for every result.
Pilot: shadow, then assisted
1
Shadow
The agent runs on real work alongside the people doing it. Its output is compared, not acted on.
2
Assisted
The agent prepares the work and a person approves it before anything happens. Anything customer-facing stays draft-with-approval.
The release decision
Release when quality, controls, acceptance and operation all pass. Each role has its part:
If anything fails, hold the release, fix what’s blocking it and run the acceptance tests again.
Skills built by business users
Skills follow a lighter version of the same discipline:- The builder approves the outline before the package is written.
- The package is tested with ordinary cases, missing inputs, conflicting evidence and requests it should refuse, and the builder reviews that evidence before installing it.
- A small review group decides what’s safe to share and names an owner.
- Every shared skill records its owner, version and evaluation set.
Where to go next
Run in production and adopt
What has to be true once it’s live.
Build an agent with the Within ADK
The full sequence from objective to deployment.
Need help? Use the in-app chat — click the chat bubble in the bottom-right corner of Within (staffed 24×5) — or email support@within.ai.