Autonomous workflows, not agent theatre
Most AI workflow projects fail because they try to be magical. The ones that survive are mostly deterministic plumbing with a small, well-placed model.
System summary
Field note on building autonomous workflows with state machines and targeted model calls instead of fragile multi-agent demos.
A workflow system is autonomous when it can run for a week without anyone tapping a button. It does not become more autonomous by adding more agents. It usually becomes less autonomous, because every additional model is one more thing that can drift, hallucinate or stall.
What survives production
- A clear state machine. Steps, inputs, outputs, retries, dead-letter.
- One model call per decision, with a deterministic fallback when confidence is low.
- A human override path that is faster than restarting the run.
- Observability that tells you which step is slow before the customer does.
Everything else is theatre. Pretty diagrams, recursive agents arguing with themselves, autonomous planners that need a human to babysit. We have stripped enough of these systems for parts to be confident about that.
The model is the most expensive and least predictable part of your stack. Use it like one.
Backend, data flows, integrations.