How the math works
Within each task
Parallelism doesn’t stop at the task level. Inside each skill run, independent tool calls fire simultaneously:Concurrency controls
Parallelism is managed at two levels to protect against rate limit cascades:
If you invoke a third skill while two are running, it queues and starts the moment a slot opens.
Rate limit handling: On a 429 response, Ultron reads the
retry-after header, falls back to exponential backoff (1s → 2s → 4s), and retries up to 3 times before surfacing an error.