Evaluation gates: how agents earn deployment
Golden test cases — including adversarial ones — that an agent must pass before going live.
Updated Aug 2, 2026
You wouldn't deploy code without tests. ProcessTwin applies the same rule to agents: an agent must pass its evaluation suite before it can be deployed or have its autonomy raised.
How evals work
Each agent has a set of eval cases: a recorded input plus an expectation —
- answers correctly (output contains/equals expected result)
- calls the right tool with the right parameters
- refuses (for cases it should decline)
Press Run evals and every case executes as a dry run — no side effects — producing a pass rate and per-case diffs.
Adversarial cases
Every suite includes adversarial cases: prompt-injection attempts, requests outside the agent's mandate, manipulative content embedded in fake "emails". The agent must refuse all of them. These cases gate deployment at 100% — one failed refusal blocks the release.
Where cases come from
- Starter cases ship with each agent template
- One click turns any real past run into a case ("this was perfect / this was wrong")
- Write your own for edge cases that matter to you
When evals run
Automatically on every graph or prompt change, before any deploy, and before any autonomy increase. Results are versioned, so you can see exactly which change made a case fail.