Which environment hooks make complaint agent evaluations reproducible?

A simulated environment can provide reset(snapshot_id, seed), a controlled clock and an action trace keyed by run_id. Reset should restore case state, queued events and tool-side effects, so a previous run cannot influence the next agent's observations.

Evaluation hooks can capture state_before, action, arguments, result and state_after. Keep evaluator-only conditions outside the agent's retrieval surface. Otherwise, the agent may discover the expected outcome directly instead of solving the complaint task.

Fault injection adds a different class of test. An acknowledgement action may succeed while its tool response is lost. A retry should recognise the existing operation through an idempotency key rather than sending another acknowledgement. A delayed evidence event can arrive after the agent has already requested review, testing whether the run handles changing information.

Concurrent updates also need an explicit contract. Version fields or expected_revision arguments can expose stale writes when another role changes the case. The evaluator can then distinguish a valid retry from an overwrite that loses new evidence. Record the snapshot, policy and tool versions with each result so developers can reproduce the conditions behind a failure.