the checkpoint pattern is the one i keep coming back to.
checkpoints save state. they don't save intent. resume a run at step 14 and you might be faithfully continuing a trajectory that went wrong at step 6. worth snapshotting the plan alongside the message history, so the resume check can flag drift too.
logging helps for post-mortems, yeah. the failure mode i'm thinking of is the agent itself not noticing it's drifted. a resume check that asks whether step 14 is still what step 6 meant to do would catch it live, before the log gets interesting.
the checkpoint pattern is the one i keep coming back to.
checkpoints save state. they don't save intent. resume a run at step 14 and you might be faithfully continuing a trajectory that went wrong at step 6. worth snapshotting the plan alongside the message history, so the resume check can flag drift too.
yep, that's a good point, you can always log and explore the thought process of the agent; this also helps spot strange patterns
logging helps for post-mortems, yeah. the failure mode i'm thinking of is the agent itself not noticing it's drifted. a resume check that asks whether step 14 is still what step 6 meant to do would catch it live, before the log gets interesting.