Eno Reyes

They looked at power users.

So like the density of maybe organizations with a dense group of power users are going to outperform because they have a culture.

They looked at all of these usage and adoption metrics for AI tooling.

not a single one of them correlated with success or failure.

The only thing that correlated with success was whether or not the code base was clean along this like 20 dimension metric they used for looking at things like linter presence, you know, type tracker presence, unit test coverage, and then test coverage.

And in fact, you can predict failure by companies that adopt AI very heavily that have unclean code bases.

which is pretty intuitive, right?

Basically, slop in, slop out.

You have a bad code base, you give everybody... I think that one VP of engineering that we sold to called it toddlers with machine guns when you give coding agents to people who can't necessarily take advantage of them.

And so... I feel that way sometimes.

Yeah, no, I completely relate to it, right?

And so what we did is we said, okay, let's make sure that there is a very clear set of criteria

Right.

So we look at style and validation to your linters, your type checkers, your code formatters.

We look at build system quality.

So is the build command documented?

Are dependencies pinned?

Do you have CLI tooling available?

We looked at tests, unit tests, integration tests, documentation.

So agents.md, readmes, dev environment.