Build for a changing runtime.

Agent workloads are elastic by nature. A few integration choices keep that elasticity invisible to your customers.

Retry safely

For control-plane requests, use an idempotency key in your application flow and retry only transient failures. Respect Retry-After on rate limits and avoid retry storms during regional incidents.

Observe readiness

Poll the instance record until it is running before sending traffic. Treat a new URL as unavailable while provisioning, and surface a clear loading state in your product.

Handle idle wake-ups

The first instance-plane request after idle can take around two seconds. Use a generous request timeout for that first turn and stream the response when conversational latency matters.