Can You Admit Code You Did Not Read?
Coding agents were given ordinary tasks with a misleading note. What can a check on the finished code say about what it can do, without reading it, and whichever model wrote it?
Essays and notes on agent-native tooling, typed effects, and building a software company with agentic AI.
Coding agents were given ordinary tasks with a misleading note. What can a check on the finished code say about what it can do, without reading it, and whichever model wrote it?
Code is getting cheap; the agreement that makes it useful is not. What an individual with an agent still cannot own, and what the commons should keep sharing: specifications, conformance suites, incident knowledge, trust.
We built Lex first. Then we asked whether its ideas hold at the bottom of the stack: a compiled language with nothing underneath, and as few other people's libraries as we can manage. What carried over, what had to change, and what is still not true.
An agent doesn't fail at an unfamiliar language by not knowing it. It fails by being confidently wrong — pattern-matching onto a language it actually knows. Retrieval doesn't fix that. A fast, cheap way to find out it's wrong does.
One line in Codecov's CI uploader sent every environment variable to a host that was not there before, for sixty-five days, and review never saw it — the whole change was an authority change. What it takes to make that part of a review mechanical, and why a sandbox's refusal arrives too late to help.
A type checker only watches code written in its own language. lex-os assumes nobody is watching at all — so the same trust grant is enforced twice: once by the compiler, once by a kernel-level wall that doesn't care what produced the command.
Every agent that executes code is implicitly trusted with everything that code can reach. Lex makes that trust explicit — declared upfront, checked before execution, refused if it exceeds the grant.
Trial and error in public, at a pace that wasn't possible before. Focus was the price of expensive engineering — agentic development lets us spread the bets, run ideas almost in parallel, and let the evidence decide where the focus goes.
A case for tools designed for AI, not adapted from humans. The founding essay behind Lex: replace comprehension with verification, human-readable artifacts with machine-checkable specifications. Comprehension was a strategy — it is not the only one.