Agents draft. Humans decide.

I work with a team of AI agents every day. This week I tidied them up, and the biggest gains didn’t come from smarter models.

They came from management basics.

A clear job for each agent: what it owns, what it refuses, and where it must stop and ask.

A hard rule: nothing goes out without my approval. An email, a post, a payment, a delete. Every one waits for a yes from me. It costs about ten seconds. It’s also the reason I trust them with real work.

Quiet routines. An agent that finds nothing new should say nothing. Since mine stopped reporting “all clear”, I actually read the alerts that do come through.

A source next to every claim. When an agent doesn’t know, I want it to say “unknown”. A confident guess is worse than an honest gap.

None of this is about better AI. It’s ordinary good management, applied to software. And it’s the same thinking behind everything I’m building: AI that does the work, and people who make the calls.

I’m building human-centric agentic tools. I’ll keep writing about the lessons as I go.