Deploying AI Agents: Containers, Orchestration, and Scaling the Loop

The article explains that AI agents differ from web services and require deployment strategies based on their longest step, not average. It recommends stateless HTTP endpoints for under 30s, WebSocket/SSE for 30s-5m, and queue+worker for longer tasks. Key advice includes pinning dependencies, running as non-root, and avoiding holding HTTP requests open for extended periods.