Replay Without Regret

An agent pipeline runs on a durable workflow engine that rebuilds state by replaying the recorded event history after a process restart. A developer calls the LLM and reads the wall clock directly inside the workflow function. What goes wrong?

The LLM call blocks the engine's event loop, so replay stalls until the outbound request finally times out

Replay re-executes those calls, gets different values, and the workflow diverges from its recorded history

Replay skips all external calls, so the workflow resumes with those variables left unset and crashes on first use

The engine cannot serialize the model response, so the workflow fails when it tries to persist the event