Summary
Peter Yang sits down with Nan Yu and Jacob Shumway, the team behind Linear’s production AI agent, to unpack five rules for building agents that actually work at scale. Using Linear’s own agent as a concrete case study, the conversation covers the full arc from initial internal memo to a shipped product that autonomously creates issues, reads Slack, scans codebases, and opens pull requests — with one demo completing an issue and PR in six minutes.
The discussion goes deep on architectural tradeoffs: why giving agents less context up front often outperforms exhaustive prompting, how Linear built a lightweight router to direct 80% of requests to a highly optimized subprompt for issue creation, and when to use a smaller model versus the largest available. The team describes their skill architecture as a key unlock that dramatically improved agent quality, and emphasizes that evals and defined success criteria must precede any model downgrade for cost optimization.
Nan and Jacob also address the compounding value of tool access — arguing that Slack integration alone is weak, but combining it with codebase access and project context creates an agent that understands existing work rather than duplicating it. The episode is a practical guide for product engineers building LLM-powered agents in production, with lessons drawn directly from Linear’s iterative development process.
📺 Source: Peter Yang · Published August 09, 2026
🏷️ Format: Interview
