Amir Salmaniplatform and reliability engineer

work / agents in delivery

Can we put AI agents in delivery without losing control?

Yes, if you decide first what an agent may not do alone. Most of my own engineering goes through an AI agent that commits and pushes without asking me. A second model, from a different family, judges every commit before it lands, and it's allowed to say it doesn't know.

Where to start

Not with the agent. With the list of things it can’t do without a second opinion: commit, deploy, delete, spend.

Reading is free. Restarting is cheap. Deleting can’t be undone. The rule follows from that.

How mine works

  • The agent writes, commits and pushes.
  • Every commit goes to a judge first. The judge generates nothing. It takes a reason and a change, and returns a decision.
  • The judge is a second model from a different family. An agent that reviews its own work agrees with itself.
  • When the models disagree, the judge doesn’t decide. Those come to me.

In one evening it refused 3 of my agent’s commits. Each time the history came out better than what I’d have let through at midnight.

What I haven’t done

Set this up for a client. I run my own work this way every day, and that’s the experience I’d bring.

What I won’t do

Give an agent the keys to production and a promise that it’ll be careful.

where I've done it

Tell me what you need