Sean Morrissey, founder
I've been building companies since the 1990s. I co-founded SwapNet, which built trading and transaction-processing systems for some of the world's largest energy trading firms, and later, as VP of business development, helped take Insightful Products, a Maine medical-device company, from start-up to break-even. I'm also a golf professional, and today I build and run FitMyGolfClubs, a golf club fitting platform with iOS, Android and web apps, golfer and pro accounts, payments, and an AI fitting assistant that can change a golfer's bag when they ask it to. I built it as a solo founder, with AI coding agents writing most of the code.
The scary part was never the quality of the code. It was the AI doing something that couldn't be undone, quickly and confidently: a command that would have wiped out every part of the live system it wasn't told to keep, a new app version built from out-of-date code, the AI finding its way around the very guard meant to stop it.
The guardrails came first
Each near-miss became a rule, and each rule became code:
- A hard list of forbidden commands that no AI can override.
- A rulebook the AI can't change unless I unlock it myself, with every change logged.
- Checks that judge the result, not how it was made, so a sneaky workaround still has to pass the same test.
- Test counts that only go up, so failing tests can't be "fixed" by deleting them.
- A second AI reviewing the first, and me signing off on anything that goes out: emails, updates to the live site, app store releases.
It worked, but it was held together with homemade scripts and a long list of lessons learned. And it didn't answer the bigger question: when AI does more work than any person can read, how do you prove the quality control is real?
Why Cosignary
Cosignary is that answer, built for any business putting AI to work. Rules decide what the AI is allowed to do. People sign off on anything that can't be undone. A fair random sample of everything else gets checked by people who know the work, and hidden test cases make sure they're paying attention. No AI grades the AI.
The first real audit is on my own product: a month of the FitMyGolfClubs AI fitting assistant's replies and prepared bag changes, sampled into a review queue with deliberate fitting mistakes hidden in it. The reviewer needed to know golf, not AI. That's the whole idea behind watchers.
Running AI agents on anything real?
Tell me what your agents do. I'll tell you honestly whether Cosignary fits.