Whenever you're ready, walk me through how you'd approach this.
Design a scalable infrastructure for deploying hundreds of concurrent, autonomous agents.
Explain how you would coordinate agent lifecycles, model inference, tool execution, memory, isolation, observability, and recovery under realistic production load. Treat latency, cost, safety, model failure, state consistency, and training-serving skew as first-class design concerns.