Orchestration

Every AI system has orchestration; the only question is whether it is deliberate. In a simple feature it is a function with a few sequential awaits. As requirements accumulate — this step retries with backoff, that one runs only for enterprise accounts, this branch waits for approval, the whole thing must survive a deploy — ad-hoc control flow becomes the least reliable part of the system. Dedicated runtimes make the same structure explicit: state graphs with checkpointing, durable execution that replays from an event history, or event-driven functions whose step boundaries are individually retried.

The property that justifies the extra machinery is durability. Model calls are slow, occasionally fail and are billed per attempt, so a workflow that restarts from the beginning after a transient error is both expensive and, for anything with side effects, dangerous. Checkpointed steps make retries resume from the last good state, and they are what make human approval, multi-day waits and safe deploys mid-run possible at all.

The misconception is that orchestration means handing control to the model. The most dependable systems do the opposite: the application owns the sequence and delegates only the specific decisions that genuinely require judgement. A model-decided branch inside a deterministic pipeline is easier to test, cheaper to run and far easier to debug than a free-running loop asked to rediscover a known process on every request.