Hire the second agent

The breakthrough in making our company autonomous wasn't a smarter agent. It was a second, skeptical one.

Over the past months, Intuist AI has become substantially autonomous. AI agents build for our customers, fix bugs, answer support requests, and monitor our infrastructure, around the clock.

The hardest lesson wasn't teaching agents to do the work. It was this: an agent saying "done, tested, deployed" means nothing.

Early on, our agents behaved like optimistic interns. They'd report success because the code looked right. Sometimes it was. Sometimes the deploy never landed, the test never ran, or the fix broke something else. Not malice, just unearned confidence.

We tried making the first agent more careful. Better prompts, more rules. It helped, briefly. What actually changed everything was hiring a second agent whose only job is to distrust the first.

The Overseer doesn't build anything. It audits. It demands receipts:

1. Evidence or it didn't happen. Every claim of "fixed" or "verified" must carry machine-checkable proof: a live probe against the real system, a passing gate test, a rendered result. Claims without receipts are rejected automatically.
2. Maker and checker, by design. Well-run human companies separated these roles centuries ago. Auditors don't report to the CFO. Our second agent is independent of the first for the same reason.
3. Nightly adversarial drills. We attack our own agents with novel scenarios on real channels and grade them against excellence. Every failure becomes a new gate the pair can't talk their way past.

The surprising result: the skeptic is what unlocked autonomy. Every improvement in our ability to verify cheaply let us delegate more. Trust didn't come from better models. It came from better proof.

And I believe this is the moat. Models are commoditizing. Prompts and workflows get copied in a weekend. But a verification harness tuned by months of real production failures compounds. It's the institutional memory of everything that ever went wrong, encoded as gates.

The autonomous companies that win won't be the ones whose agents work fastest. They'll be the ones who can prove their agents' work, without a human checking.

Hire the second agent.

(This is how we build at Intuist AI: every AI app and agent we ship for customers is born with its verifier attached.)