Building agentic applications requires moving beyond simple web request handlers to address their long-running, stateful, and non-deterministic nature. Naive approaches, where an agent's loop is tied to a single HTTP request, are fragile in production.

Production systems employ three core patterns. First, decouple agent runs from web requests by enqueuing work for durable execution. A request creates a run record and enqueues the job, returning a run ID immediately. A separate worker executes the job, with progress logged in a database. This pattern uses queues as durable buffers, suitable for tasks longer than a request, independent tasks, and when worker scaling is needed.

Second, ensure reliability through idempotency and compensation. Idempotency handles steps that might run multiple times, while compensation addresses partial failures by reversing completed actions. This is crucial for agents interacting with external services.

Third, leverage workflow engines for orchestration. Workflow engines manage the state and decision logic of an agent's run, storing history and ensuring consistent decision-making even after crashes. This separates decision logic (coordinator) from execution (steps), enabling robust recovery and replay.

These patterns transform brittle agent scripts into scalable and resilient production systems by managing unpredictability and preserving progress.