Test

Validate the agent before anyone depends on it.

Test is where you find out whether the agent actually does what you intended — before it reaches real users.

Testing an agent is not like testing ordinary software. The same input can produce different output, so a single successful run proves very little. What you need is a scored dataset of representative cases, run repeatedly as you change instructions, skills, and behavior settings, so you can see whether a change improved the agent or merely moved the problem.

• Evals — Test the agent against scored datasets before release.

You can also exercise the agent interactively with Try Agent in the builder, and open the execution trace to see how it interpreted your instructions, which skills it called, and how long each step took. Use interactive testing to explore and evals to prove.

---

When the agent performs on your test cases, move on to Deploy. After release, Monitor tells you how it behaves on traffic your test set never anticipated.