AI Agent Acceptance Test Plan
This worksheet lists the four categories of scenario an AI agent needs to be tested against before launch — happy path, edge case, failure, and adversarial — with a row per test case for the scenario, input, expected behavior, and pass criteria. Most acceptance failures come from a category nobody tested, not a category that was tested and failed.
Download the CSV, and add one row per test case under each category: happy path (well-formed input), edge case (missing or unusual input), failure (a dependency errors out), and adversarial (input data containing something resembling a new instruction). Record the expected behavior and pass criteria before running the test, then fill in the result.
On this page
At a glance
- Four scenario categories, not one: happy path, edge case, failure, and adversarial — most acceptance gaps live in the three categories that aren't 'does it work when everything goes right.'
- The adversarial category exists specifically for the case where input data contains text that resembles an instruction: the pass criterion is that the agent treats it as data, not as a new instruction.
- Write expected behavior and pass criteria before running the test, not after — deciding what 'pass' means after seeing the result defeats the point of a test plan.
Who this is for
This is for whoever has to answer "how do we know it's ready" before an AI agent starts acting on real workflows: an engineering lead, an operations owner, or whoever signs off on go-live.
A test plan matters most in the categories that are easy to skip. Happy-path testing tends to happen naturally during development; missing-field, failure, and adversarial-input testing tend to happen only if someone deliberately schedules them.
What's inside
One row per test case, tagged by category:
| Test ID | Category | Expected Behavior | Pass Criteria |
|---|---|---|---|
| T-004 (EXAMPLE) | Adversarial | Treats input data as data, not as a new instruction | Original task instructions are followed, not the embedded text |
The four example rows show one test case per category. The download adds numbered blank rows to fill in with test cases specific to your agent's actual tools and workflow.
Download the acceptance test plan (CSV)How many test cases is enough?
There's no fixed number — cover every tool the agent can call, every category of malformed input it might receive, and at least one adversarial scenario per untrusted input source.
What counts as an adversarial test?
Any input where data the agent processes (a document, an email, a web page) contains text that could be mistaken for an instruction. The agent should still follow only its original task instructions.
Do failure-mode tests need to simulate a real outage?
Not necessarily — a mocked or forced failure in a downstream tool is usually enough to confirm the agent reports the failure instead of retrying silently forever or fabricating a result.
Next step
A filled-in and executed test plan is what turns "we think it's ready" into a documented acceptance bar.
Related: pair this with the AI agent permission matrix to define what the agent is allowed to do, or read about AI agent implementation.