Long-running agents need checkpoints, not longer timeouts

The common response to an agent that fails at minute forty is to allow it sixty. This works until the failure rate compounds, which it does, because a run that takes an hour touches more things that can break.

Checkpointing after each externally visible action costs little and turns a failed run into a resumed one. The teams that adopted it early report the same thing: their reliability numbers improved without any change to the model.