Module 1 · Simulated worlds · Lesson 1
Stateful environments and end-state grading
Why multi-turn agents need an environment rather than tool stubs, and how to grade the world an agent leaves behind with gold-action replay, state hashes, targeted assertions and collateral-damage checks.
By the end of this lesson you can
- Model an agent evaluation as agent, environment and policy, and say which state and rules each one owns
- Derive a target end state by replaying a reference trajectory on a fresh environment, and compare end states by hash
- Split a failed end state into missing changes and collateral damage with a cell-level diff against the initial state
- Combine database, communication and assertion checks into a product reward, and explain why benchmarks rarely grade action sequences
- Design an environment whose writes are deterministic and canonical enough to hash