Introducing Code Reality Labs

For the last couple of years I’ve been removing the same bug over and over. Not from a single codebase — from the way AI agents work.

An agent is only ever as good as what it knows. Give a brilliant model the wrong picture of your system and it will hand you confident, fluent, plausible, wrong code — and do it fast. The model was never the bottleneck. The bottleneck is context: what the agent believes about your code, your history, and your intent. When that’s a guess, everything downstream is a guess wearing a suit.

Code Reality Labs exists for one reason: to make the context AI runs on real. Correct, compact, persistent, verifiable. Not a vibe — ground truth.

It started as irritation, not a business plan

I didn’t sit down to build five products. I wanted my own agents to stop guessing, and every time I fixed one reason they guessed, the next one underneath it became impossible to ignore:

• They had no reliable model of the code in front of them. → TheAuditor
• I couldn’t prove a code-analysis tool’s claims were any good. → BenchProctor
• The client wrapping the model wasted the context I’d just built, drowning facts in prose. → Warden
• Even perfect knowledge of today’s code is amnesiac the next morning. → Curator
• Long-running work across all of it had no supervisor. → Arbiter

Five products. Each one earns its keep on its own, in its own domain. But look at the list and the through-line is obvious: I kept removing the reasons agents are stupid, and the reasons formed a stack.

One standard, five front doors

The thing that makes these cohere isn’t a shared codebase or a forced bundle. It’s a single question: what would AI agents need if their context had to be correct, compact, persistent, and actionable? Answer that honestly and you get exactly these five shapes.

That’s why they live under one roof. Not so you have to buy all of them — you don’t, and you shouldn’t have to. Each is built to be the best tool in its lane standing completely alone. It’s so the standard behind them stays consistent: the same refusal to let an agent act on a guess.

Two of them are open source — Warden (MIT) and BenchProctor (Apache) — because some things you should be able to read, run, and check for yourself. The benchmark most of all: a measurement you have to take on faith isn’t a measurement.

What this blog is

This is the company’s voice, not a product’s. Here we’ll write about the through-line — where the stack is going, the principles behind it, and how the pieces fit. Each product keeps its own blog for the deep, specific work; this is the altitude above them.

We called it Code Reality Labs because that’s the entire job, said plainly: keep AI tied to the reality of your code, not its imagination.

Welcome. Let’s stop letting agents guess.

— John, Founder · Code Reality Labs