TL;DR
The short version
Warp CEO Zach Lloyd shows Claire Vo a cloud software factory (Wilson): public Slack kickoff → Linear → GitHub PR → computer-use QA video, plus manager metrics and LLM-as-judge self-improvement.
On How I AI (21 Sep 2026), kickoff-to-PR is ~35 minutes while PR-to-first-human-review is ~3.5 hours — generation got cheap; humans are still the bottleneck.
Built on Claire Vo’s How I AI walkthrough with Warp CEO Zach Lloyd (21 Sep 2026).
Builder view: work starts in public
Lloyd’s builder flow is not “open laptop, prompt a coding agent, push.” Someone tags Wilson in a public Slack channel. Wilson treats the ask as factory input: triage, open a Linear issue, implement, open a GitHub PR, then prove the change with computer-use verification — in the demo, a video of the feature being exercised.
Local agents stop at “I think this works.” The factory tries to ship ticket + code + review surface + evidence together, where the team can see the thread. Triggers are not Slack-only: Sentry crashes can enter the same line; work can start in Linear or GitHub.
Manager view: human interactions per PR
When every engineer has a private laptop agent, leaders cannot answer whether automation, cost, or quality is improving. Warp’s control plane adds the usual velocity and cost views — and one Vo hadn’t seen: human interactions per PR.
Humans are a little bit of the bottleneck in terms of production and software. They're also the creative force.
The metric rolls up Slack reprompts, Linear comments, and review corrections — how often humans stop the line. Vo’s hard time lands on the same chart: kickoff to PR about 35 minutes; PR to first human review about 3.5 hours (Warp’s live dashboard, self-reported). Warp still human-reviews every PR; they did let the prompt author review agent code instead of forcing a second reviewer.
Score, observe, change the factory
Because runs are centralized, Warp scores them. A cheaper LLM-as-judge looks across aggregate runs for failure modes — redundant tests was the live example. Then an observer agent studies a batch of failed runs (~20–25) and proposes updates to the factory definition.
What is a software factory? For us at least, it's an actual noun… repos, MCP servers, configuration and then a bunch of agents… and it's all defined in code.
Factories-as-code lets you freeze a version, replay past tasks across model configs, and build a cost–quality Pareto on your work — then route. Warp’s public product page matches the demo shape (warp.dev/factories; launch post).
Chaos control at thousands of PRs
Vo asks how you keep arms around the team when the factory hums at over 2,000 PRs in a month (self-reported). Lloyd’s answer is almost boring: dogfood hard, let power users work in public so the average uplevels by watching, and keep user interviews and design jams so speed doesn’t outrun the right stories.
Put agent work somewhere the team can see it. Measure human touches per shipped PR, not only tokens or PR count. Treat review wait as the bottleneck when generation is already cheap. Score runs on your failure modes and require a sample before you “self-improve.” Keep a human product process when throughput explodes.
