dropping history replay would remove so much operational pain from durable workflows. how do you reconstruct in-flight state after a crash, snapshotting or something event sourced
Checkpointing the live continuation shifts recovery cost toward retained state rather than lifetime event count, which seems especially useful for agents with many completed tool calls.
4 comments