Linear, valued at $1.25B, built an internal AI agent and two of its architects, Nan Yu and Jacob Shumway, explain exactly how they did it. The episode traces the full arc from the initial memo to production deployment, covering the specific engineering decisions that made Linear Agent reliable enough to use in a real company.
The technical substance is what makes this worth watching. The discussion covers how to give agents the right tools to retrieve context dynamically, and how to use evaluations to systematically improve reliability over time. These are not abstract principles. They are the actual methods Linear used on a product serving a large-scale user base.
The five rules framing this conversation are the starting point, not the conclusion. The real value is in the process Nan and Jacob describe between the rules: how evals are structured, how context retrieval is designed, and where production agent builds tend to break down. If you are building agents beyond the demo stage, this is the specific, practitioner-level detail that is hard to find elsewhere.
[WATCH ON YOUTUBE →]