Linear Agent dynamically loads task-specific skills in production

Linear Agent runs an LLM in a tool-calling loop, loading task-specific skills and context from systems such as Slack, Linear, and a codebase to turn a Slack discussion into a Linear issue and a pull request.

• Linear’s first prototype called an LLM directly from the frontend, exposed Linear command-menu actions as tools, and was initially released internally through Slack app mentions without an announcement.
• The production architecture replaced a large always-available action set with dynamically loaded skills: the agent calls a tool to load the skills, tools, and instructions required for a request, rather than receiving all context or the full GraphQL schema upfront.
• In the Slack example, Linear Agent read the thread, created and assigned a Linear issue, delegated work to itself, and produced a pull request in 6 minutes; the issue remained assigned to Jacob Shumway’s to-do backlog.

Uses Codeex, claw code, and open code to research and optimize Polymarket order execution latency, aiming to obtain early FIFO queue positions for low-priced resting orders rather than making prediction-based trades.

• On a five-minute Bitcoin up/down market test, five shares were bought at 5 cents and resolved for a $5 payout, described as a 10,000% profit.
• Placed an order for five “Yes” shares at 1 cent on “US recession by the end of 2077”; the terminal estimated 12 shares ahead of the order and 1.1 million shares behind it.
• Execution optimization components listed were Polymarket WebSockets instead of polling, precomputing parameters and credentials, VPS hosting, persistent HTTP connection reuse, and profiling signing and execution latency.

Adam Ward’s hiring process uses three steps—scoping the required skills and experiences, mapping a finite target list of roughly 50 qualified people, and maintaining personalized contact until they enter the process or join.

• Ward describes the “funnel of doom” as contacting 100 people and receiving replies from about 20; he says those 20 respondents are not necessarily the top 20 candidates, but the people reached at a moment when they were receptive.
• For market mapping, recruiters ask trusted contacts attribute-specific questions such as which product engineer was most collaborative with designers or best at translating a framework into a product, then triangulate repeated names into a target list.
• Cursor uses role-specific work trials during longer onsite experiences: engineering candidates work side-by-side on projects, go-to-market candidates complete a customer challenge with the hiring manager, and product candidates complete a project; interview debriefs usually occur the same day or the next day and are hiring-manager led rather than handled by a hiring committee.