Linear data: agent-driven teams ship 3x the pull requests
Linear's 127k-user report lands as coding-agent teams triple their PRs, AI now writes half of all issues, and PMs and designers start shipping code.
Copy markdown
Coding-agent teams ship 3x the PRs
Across 127,000+ paid users, Linear found teams using coding agents jumped from 21 to 65 merged pull requests a week — while teams without them barely moved, 8 to 10. That gap is now the cleanest ROI signal for putting an agent in your workflow.
AI now writes half your issues
Nearly 50% of all issues created in Linear are now authored by AI, up from fewer than 1 in 1,000 two years ago. If you triage a backlog, expect most tickets to arrive machine-drafted — tune your templates, labels and triage rules for it.
Designers and PMs are shipping code
Designers attaching PRs went 1% to 8%, PMs 3% to 10%, and founders 11% to 23% in six months, with AI adoption more than doubling across every role. Yet teams spend more time on product work, not less — a Jevons effect where cheaper output just means more of it.
Map any Hugging Face model, free
modelmap.cc renders interactive, animated architecture diagrams for any Hugging Face model — a fast way to learn or debug an unfamiliar transformer's internals before you fine-tune or port it. It front-paged on Hacker News this week.
Sandbox untrusted agent code
Simon Willison walks through using smolmachines/smolvm as a lightweight VM to run untrusted Python and JavaScript from agents, spinning it up on GitHub Actions runners with /dev/kvm when local nesting isn't available. A practical pattern if your agent executes model-written code.
Elsewhere: Vercel adds Algolia and Slack ops
Algolia search is now installable straight from the Vercel Marketplace with unified billing, and Vercel for Slack entered public beta — investigate production incidents and run remediations without leaving a Slack thread.