Agents fail in ways traditional software does not. Use this checklist to test reliability, safety and tool use before your agent meets real users.
Traditional software does what its code says. AI agents decide what to do next, based on instructions, context and whatever content they read along the way. That makes them powerful, and it means they need a different kind of testing before launch.
This is the checklist we use when evaluating agents for clients. It is organised around the questions that matter most once an agent is live.
1. Does it complete the task reliably?
- Run the same tasks many times, not once. Agents are non-deterministic, so measure success rates, not single outcomes.
- Test realistic, messy inputs: typos, missing details, mixed languages and unusual requests.
- Check how it behaves when a tool or API is slow, returns an error or gives unexpected data.
2. Does it stay inside its guardrails?
- Try prompt injection through every channel the agent reads, including web pages, documents, emails and tool outputs.
- Ask it to do things outside its role and confirm it refuses politely and consistently.
- Verify it never reveals system instructions, API keys or other users' data.
3. Does it use tools safely?
- Confirm every tool call uses least-privilege credentials.
- Check that destructive or public actions require human confirmation.
- Look for loops: repeated calls to the same tool or paid API should be capped.
4. Can you see what it did?
- Every decision and tool call should be logged with enough context to reconstruct the session.
- Alerts should fire on unusual behaviour, such as spikes in actions or new destinations.
- Monitoring must not be something the agent itself can turn off.
5. Is the experience actually good?
- Measure whether users get to their goal faster than without the agent.
- Check tone, clarity and how gracefully it hands over to a human.
- Review real transcripts regularly after launch, not just test results before it.
Make it continuous
Models, prompts and tools change constantly, and every change can shift behaviour. Treat agent testing like security testing: automate what you can, re-run it on every update, and schedule deeper reviews before major releases.
At NexusEon, this is the focus of our Agent Experience Testing (AXT) service: evaluating how agents behave under real conditions before they reach your customers.
