An AI workflow must call several downstream services. One service has a strict concurrency limit and cannot absorb the model agent burst rate. Which design best protects the dependency?

1. Increase the model context window as the only capacity control.
2. Suppress throttling metrics so the workflow appears healthy.
3. Bound parallel work and queue or schedule tool execution according to the downstream capacity.
4. Launch every tool call concurrently and retry all throttles immediately.