Infrastructure Patterns for Agentic Applications

# Summary Agents—AI systems that iteratively call models and tools to complete tasks—have fundamentally different infrastructure requirements than traditional web applications because they are long-running, stateful, and non-deterministic. The naive approach of wrapping an agent in an HTTP request handler fails in production because it couples application reliability to unpredictable execution times and lacks durability for handling crashes or retries. The article describes the Web-Queue-Worker pattern as the first production solution: instead of running agents within a request, the system should immediately enqueue the work, return a run ID to the client, and execute the agent asynchronously via a worker process that persists progress to a database.