Warp ships 2,000 pull requests a month. CEO Zach Lloyd, former principal engineer on Google Sheets, built a system called Wilson that runs a full pipeline from a Slack message through Linear ticket creation, GitHub PR, and automated QA, with minimal human input per task. The key metric Warp tracks is not velocity but human interactions per PR, a direct signal of how much the automation is actually carrying.
The architecture details are worth your time. Lloyd explains how Warp scores every agent run using LLM-as-a-judge, replays real historical tasks to build a cost-quality Pareto chart for model selection, and feeds failed runs back into the factory to self-improve its own workflows. Human code review remains the bottleneck, not generation. That constraint shapes every design decision in the system.
The second half shifts to Lloyd's personal CEO workflows using Figma MCP, Granola, and research agents, plus a live GOG CLI demo. His core reframe: stop thinking in sequential tasks and start thinking in parallel agent runs. If you are building or evaluating agentic engineering systems, the sections on scoring, cost tracking per PR, and failure-mode analysis are the ones to read in full.
[READ ORIGINAL →]