The journey to shippable AI systems: Patterns that work

This post introduces five patterns for building shippable AI systems that leverage durable execution: 1. Guarded tool calls (timeouts, retries, circuit breakers, fallbacks, cost metadata). 2. Idempotent side-effects (idempotency keys, compensating actions). 3. Human checkpoints (real approval steps, reviewer timeouts). 4. Deterministic plans, versioned prompts. 5. Cost and policy budgets. It frames these patterns as solutions to common operational issues in AI workflows such as overhead, long-running complexity, and brittle recovery, and reinforces the role of durable execution in providing state, retries, visibility, and long-running task management for these AI systems.