My Agent Never Said "I Don't Know"

The useful part

This post is about four bugs I hit while building that Agent. My front-end and back-end colleagues gave me ideas, pointed out problems, and occasionally pulled me out of rabbit holes I'd dug myself into. The product decisions, the Harness design, the debugging loop — those were mine to own.

How it works

• In an AI-native stack, the gap between "I have a product idea" and "it's running in production" has collapsed.
• A PM who can design model behavior, engineer context, and debug Harness failures doesn't need to wait for engineering bandwidth — they just ship.
• When they do need engineers, they can have a real conversation instead of throwing a spec over the wall.
• Some call it FDE — Field Development Engineer, or Founding-level engineer who sits at the product-engineering boundary.

What to take from it

Someone who owns the full loop: requirements, model behavior, Harness design, deployment. A builder who understands the product deeply enough to make the right engineering calls.

Details worth keeping

I write specs, run reviews, align stakeholders. Last year I got tired of handing things off and waiting. Contents # Bug Hallucination Type Fix 1 Recommended a gun that doesn't exist Knowledge gap → generic fill Explicitly mark absent entities in KB 2 Fabricated a player ID that looked real Tool use parameter fabrication Lock parameter source to context injection 3 Got its own ID wrong Identity inference...