Come build the yardstick.
Every engineering organization is choosing AI coding agents, and almost none of them can compare those agents on their own work. We are building the tool that lets them.
Open roles
Founding Engineer, Agent Runtime
You own the path from one request to two finished results: running two agent setups at once from the same starting point in a customer's own GitHub Actions, and capturing every step each one takes. Small differences between runs quietly bias the result, so keeping them identical matters as much as making them fast.
You have built developer tooling or build and CI infrastructure that other engineers depended on, and you have opinions about what should happen when a runner dies halfway through a task.
Applied Researcher, Measurement
You own how individual picks become a verdict: the statistics behind it, how many comparisons a decision needs, and the checks that show the test itself is working. Our method page sets out the principles.
You have run experiments on real systems where the result changed a decision, and you can explain a confidence interval to a VP of Engineering without either patronising them or hiding behind the maths.
Founding Engineer, Product
You own everything engineers and their leaders touch: the blind A/B review and the results. The hard parts are making the review fast enough that engineers use it on every task, and presenting uncertainty plainly to people who want a verdict.
You have shipped interfaces for technical users and you would rather show someone an interval than a green tick.
What the work is like
teamSmall, and staying that way for a while
You will be one of the first engineering hires. Everything is unowned, including the decisions you will have to live with for years.
customersYou will talk to them
Our customers are engineering teams deciding which coding agents to use. Every hire sits in those conversations. Nobody here is insulated from the people using this.
standardsWe report what we find
Including when a campaign finds no difference, and including when it goes against the answer a customer wanted. If that sounds like an obvious commitment rather than a costly one, we should talk.
locationRemote, US time zones
With regular time together in person. We are not trying to run a fully asynchronous company at this size.
How to apply
There is no applicant tracking system. Send a note through the contact page saying which role and what you have built that is relevant: a repository, a paper, a system you can describe. We read everything and reply either way.
If none of these three is quite you but the problem is, write anyway and say what you would want to own.
Read the method first.
It is the clearest picture of how we think, and the fastest way to work out whether you would enjoy arguing with us.