How Linear loads skills and context for its agent

In a Peter Yang interview, Linear's Nan Yu and Jacob Shumway describe moving from broad static context and a full GraphQL schema toward agents that load the relevant skills, tools, and guidance for each request, then using production feedback to expand their eval set. Why it matters: Teams can test a few workflows deeply, but users can ask an agent for anything. Loading task-specific context cuts noise, and saved failed runs make the supported flows measurable.

Try this: Choose one workflow your team already runs. Replace its large static prompt with a retrieval path for the relevant skill or record, save five failed runs as regression cases, and measure the hard requirements before widening access.