Agents need a state machine

The most common agent architecture in the wild is a happy path with extra logging.

Upload. Prompt. Tool. Answer. Screenshot.

That is a path. It is not a system.

A system has states, illegal transitions, and a story for crash recovery. It distinguishes attempted from succeeded. It does not retry an ambiguous commit. It does not treat a provider 200 as proof that money moved.

When I build agentic products, I start with the state machine:

• What is proposed versus sealed?
• Who is allowed to authorize?
• What happens if the world drifted?
• What happens if we die after the effect and before the ledger write?
• What artifact proves the chain later?

LangGraph is useful here not because graphs are fashionable, but because pause/resume and typed state force you to name the gaps. Human-in-the-loop is not a UX flourish. It is a trust boundary.

If you cannot draw the states, you are not ready to let the model touch the world.