
Agentic coding
We build coding environments around real repositories, where agents work in the terminal to write, debug and ship software, so models finish engineering work the way senior engineers do.
We study the limits of today’s models and build the training data and environments that expand what they can do.
Frontier models already know more than they can reliably use, and closing that gap is a research problem. We study where models fail and build the training data and environments that turn knowledge into dependable capability.

We build coding environments around real repositories, where agents work in the terminal to write, debug and ship software, so models finish engineering work the way senior engineers do.

We build desktop and browser environments where agents operate real software through the screen, fill in forms and complete tasks with an experienced user's judgment.

We capture economically valuable work in finance, legal and consulting, from multi-app workflows to models, memos and decks, so models deliver work an expert would sign.

We build environments where agents read papers, replicate results, run experiments and post-train models, the long-horizon work that lets AI improve AI.

We build tasks where agents create playable games, interactive 3D scenes and manufacturable parts, judged on how they look, play and work.

We build environments where agents create and extend real iOS, Android and React Native apps, held to the polish people expect from apps they use daily.

We design private evaluations grounded in the work your users do, so every checkpoint is scored on outcomes that matter and each failure points to a specific skill.
Training data, environments and evaluations for frontier labs and enterprises.