I build multi-agent AI systems, and I've run mine in production for a year.

Why an AI can't just read the invoice and be done

OCR will read a number wrong sometimes. On a money document that's the whole game, because an owner who catches one bad number stops trusting all of them. So every field the pipeline pulls carries a confidence score, and anything below the line gets routed to a person to confirm before it lands in the books. The machine does the typing. A person still signs off on the number.

Why the rules are mechanical gates, not just instructions

You can write “always verify before you delete” into an agent's instructions, and it will hold right up until the agent is under pressure and reasons its way around it. Soft discipline loses to the model's own confidence. So the load-bearing rules live in mechanical gates that run outside the agent and can actually stop it. Anything you genuinely cannot afford to have go wrong, you enforce in a layer the agent can't talk itself past.

When it's worth writing a server from scratch

Most of the time the right move is to integrate something that already exists, because the boring server someone else maintains is a server you don't have to. I wrote my own only where the thing I needed didn't exist yet. The test isn't whether I could build it. It's whether the gap actually hurts and will keep existing.

Why the human stays in the approval seat

A pipeline that publishes on its own is a pipeline you stop trusting the first time it publishes something wrong. Generation is automated end to end; the publish step waits for a person. That single gate is what makes the rest of the automation safe to run at full speed.