AI agents have advanced rapidly in software, a domain with a clear iteration loop: write, run, verify, refine. We believe that the next frontier for AI agents is in long-horizon work in complex fields such as 3D and CAD. This work represent billions and sits on top of millions of expert hours that remain entirely manual to date.

Software agents work because the data exists. Code, diffs, reviews - every step of how software gets made is public, traceable, and machine-readable. No other expert domain has this. Project files stay internal and firms ship outputs, not workflows. The causal chain of decisions that turns a brief into a finished artifact lives in the heads of senior practitioners and disappears the moment the file is saved. To make agentic expert work a reality, models will need data that captures how the work happens - not just the final artifact or generic computer-use traces, but the full causal chain of decisions behind it.

Shader is building the data foundry for agentic work. We capture native tool-use traces from vetted experts performing real production tasks inside the software they already use, starting with 3D. This will lay the foundation that vendors and labs need to ship intelligent copilots and agents.