Module 2 · Designing servers agents use well · Lesson 4
Testing and evaluating servers
Pin a server's contract with in-process protocol tests in both eras and scripted Inspector checks, then find out whether agents actually do better with it, using checkable end states, repeated runs and a paired comparison.
By the end of this lesson you can
- Write protocol-level tests that drive a server through a real client in-process, for both 2026-07-28 and 2025-era clients
- Use the Inspector CLI as a scripted check and branch on its exit codes
- Design agent evaluation tasks with end states checked over the API, and test those checks with reference solutions before running a model
- Compare two server designs with repeated runs paired by task, reporting pass rate, turns, REST calls and tokens