Warp Ships 2,000 PRs/Month via a Full Slack-to-Merge 'Factory' With a Self-Improving Harness
Summary
Warp CEO Zach Lloyd describes 'Warp Factories': Slack request -> triage -> Linear ticket -> implementation -> PR -> computer-use QA -> merge, producing a first PR in ~35 minutes. Uses 'human interactions per PR' as the true throughput metric, an AI judge scores every agent run (catching redundant tests that don't guard against regressions), and the factory improves itself by mining 20-25 failed runs into an observer-agent-proposed update to agent definitions stored as code. Replays real historical jobs against different models on a cost-quality Pareto chart instead of trusting public benchmarks.
Why this signal
A rich, concrete end-to-end pipeline example directly comparable to Ramp's Inspect and OpenAI's agentic factory this cycle — the self-improving-harness-from-failed-runs pattern and 'human interactions per PR' metric are both immediately adoptable measurement ideas for Thomas Engineering.
Original email
Lennys Newsletter · 9/21/2026
🎙️ How I AI: Meta's Muse review + How Warp ships 2,000 PRs a month with AI factories