Long-running tasks, subagents, recovery, and observability
Durable state, explicit ownership, and candidate-bound evidence allow restart, worker coordination, and integration over a moving target.
Links containing ?t= open the video at a specific second.
Video summary
The ideas to retain
A long-running task is a durable state machine, not an infinite conversation
In a short interaction, we can pretend conversation, plan, workspace, and execution are one object. That simplification breaks down over hours.
“Memory” is not one object
Agent systems use memory to mean several different things:
Context, compaction, checkpoints, and durable state are not synonyms either
Compaction reduces or transforms the context fed back into the model. Its purpose is to keep reasoning within a token budget.
Key moments
Jump directly to a section
Next step
