Remote engineering is the default. Distributed AI engineering is harder than distributed CRUD engineering, because AI work depends on tight feedback loops between prompt, eval, and review, exactly the loops time zones break. The teams that get this right treat distributed AI as an operational design problem, not a culture slogan.
What's different about AI work in distributed teams
In a deterministic codebase, an asynchronous review is fine, the diff says what it says. In an AI codebase, the diff is a prompt or an eval threshold, and the review needs to look at sample outputs, eval deltas, and the cost implications of the change. That requires real-time discussion, not just an async PR comment.
Our operating system for distributed AI pods
- Four-hour overlap windows, non-negotiable, between every pod member
- An eval dashboard everyone can read at any time, with deltas highlighted
- A weekly 'prompt review', like a code review, but for prompts and eval rubrics
- A shared incident channel where every quality regression becomes a public learning
- Quarterly in-person weeks per pod, model choice and architecture decisions are made face-to-face
The communication tax, paid up front
We over-document the things that matter, eval thresholds, prompt versions, decision logs, and under-document the things that don't. The tax of writing things down once is far smaller than the tax of mis-aligning across time zones for weeks.
If your distributed team can't say in 30 seconds why the eval pass-rate moved last week, the time zones are not your problem, your operating system is.
How we hire for it
Distributed AI work rewards engineers who can write clearly, reason about probabilistic systems, and operate without constant supervision. Those are not the same skills that win whiteboard interviews. We hire for written communication, eval literacy, and judgement under uncertainty, in that order.
What we don't compromise on
We don't compromise on the four-hour overlap, the in-person quarter, or the weekly prompt review. Cut any of them and the feedback loops break. Once they break, the eval pass-rate slides quietly for weeks before anyone notices.