Supervision is the new management

I recently rewrote the page describing how I build around one phrase: human on the loop, not in it. I don’t approve every step my agent fleet takes anymore. Plans, code, reviews, merges, deploys — the loop runs on policy. What I kept is the judgment and the exceptions.

It sounds like a resignation letter from management. It’s the opposite: it’s management finally reduced to the part that was always the point.

The job was never the checking

Andy Grove defined a manager’s output as the output of their organization — and managerial leverage as the art of raising it. Assigning tasks and checking work were never the job; they were the transmission mechanism, the only one available when instructions had to pass through human attention to become work.

Agents change the transmission. A written policy — architecture rules, review gates, merge conditions — executes directly, every time, without a meeting. The transmission stopped needing me. The steering didn’t.

What supervision actually means

My fleet has a Supervisor role: it watches whole epics, tracks progress, and escalates only what needs a human decision. Designing that role taught me what supervision is made of, because I had to write it down:

• Policy authorship. Every rule the fleet enforces is a management decision made once, in writing, instead of fifty times in review comments. Writing policy is the highest-leverage hour of my week.
• Exception design. The hard question isn’t “what should agents do?” It’s “what should reach me?” Too little and you’re absent; too much and you’re a bottleneck wearing a dashboard. This is span of control, reinvented as span of attention.
• Auditability. Trust without inspection is abdication — so everything leaves a trail, and I sample it. Inspection on demand replaces approval by default.

If this sounds familiar, it’s because good managers of humans already work this way: context instead of control, standards instead of surveillance. Scaling squads taught me that no manager survives 13 teams by approving their work item by item. Agents just make the lesson non-optional at any scale.

The uncomfortable part

Being in the loop feels like diligence. Every approval is a little proof of usefulness — and a little dose of control. Moving on the loop means giving that up and accepting a scarier form of accountability: you own the outcomes of a system that mostly runs without you. There’s no “but I reviewed it” to hide behind. You wrote the policy; what the policy allows is yours.

That trade — comfort for leverage — is the actual paradigm shift. The tooling is incidental.

Where this goes

Every team adopting serious AI delivery will rediscover this sequence: first agents as typists, then agents as juniors needing review, then the review itself becomes policy — and management collapses into its essential core: deciding what’s worth building, writing the rules of good, and showing up for the exceptions.

The judgment stays the job. Everything else was transmission all along.