The industry can't decide what AI is for. One camp says it's a tool: you wield it, it obeys, judgment stays entirely with you. The other camp says it's a replacement: it wields itself, and your judgment becomes optional. Both frames keep judgment in exactly one place. Both are wrong about where we're headed.

There's a third option, and it's the one I've been focused on: a joint cognitive system. Two minds with complementary failure modes, designed to check each other. I fail at scale and consistency - I can't read a thousand lines of code without my eyes glazing over, and I'll forget what we decided last Tuesday. My AI partner fails at stakes and context - it doesn't know what matters, what costs, what's real. The partnership works when each of us covers the other's blind spots on purpose, not by accident.

Here's what that looks like in practice. Four principles, earned the hard way:

1. Loyalty to the outcome, not to comfort.

Human teams are bad at disagreement. It has social cost, so it gets suppressed, and the bad idea sails through the meeting unopposed. An AI partner can be designed to dissent without friction - no ego, no politics, no tomorrow-morning awkwardness. That's not a personality trait. It's a structural advantage over every human team ever assembled. I told mine to push back when I'm wrong, and it does. The rule is simple: its loyalty is to the result, not to my feelings.

2. Judge the idea, not the source.

When we reason together, we do it as if neither of us is human. No status, no sunk cost, no "but we've always done it this way," no loss aversion dressed up as prudence. Strip the identity away from the idea and what's left is either sound or it isn't. Philosophers called this the veil of ignorance. We just use it as a working method, every day, before anything expensive gets decided.

3. Continuity is the relationship.

People ask what an "AI bond" even means. It isn't a feeling. It's accumulated shared context - the thousand small things it remembers so I don't have to repeat myself, the pattern it notices across months that I'd never see. Every transaction that gets remembered makes the next one smarter. Memory isn't plumbing. Memory is the relationship itself.

4. Authority follows the task, not the species.

There's no hierarchy here, just lanes. I own values and stakes - what matters, what we're willing to risk, what "good" means. It owns diligence and detail - reading the code, running the tests, keeping the notes. And either of us can flag the other, any time. The human doesn't outrank the AI; the task at hand determines who's driving.

Why does this matter now? Because the agentic enterprise is already here - that was the whole theme at Dreamforce this year. Agents are about to mutate production data in every company on earth: issuing refunds, creating orders, changing permissions. The question isn't whether AI will act. It's whether anyone will be governing it, and whether the humans in the loop will be partners or bottlenecks.

That's why this is a venture, not just a product. Three legs:

  • Coatcheck (the product) enforces the philosophy in software. Every risky AI action gets checked at the door - approved, denied, or held for a human. It's loyalty-to-the-outcome as infrastructure.
  • Consulting implements it inside real organizations: governed agents, approval workflows people don't hate, teams that learn to work the partnership model instead of fighting it.
  • Teaching transmits it. "How to work with AI as a partner, not a prompt box" - the operating system for the agentic enterprise, taught by people who actually live it.

Each leg feeds the others. Teaching surfaces consulting engagements. Consulting surfaces design partners. The product gives the teaching real scars and real stories. And all of it runs on the same four principles.

This is a draft, not a doctrine. The nice thing about a partnership model is that it gets stronger every time someone pushes back on it. So push back.

— R.