How Software Factories Improve Themselves — Suraj Gupta, Warp

There's a lot of talk about self-improving agents and about software factories, but much less about how a factory itself gets better over time. Suraj Gupta, who leads harness development at Warp, shows three practical ways to do it, with demos from Warp's own open-source factory. The first is skills that improve themselves. An outer-loop agent watches the triage agent's runs and feedback, then opens a PR to update its skill, so every change is tracked in Git and reviewed by a human. The second is persistent memory: a versioned, traceable store of facts, so a Sentry agent doesn't have to rediscover a root cause it already found. It works with any harness, including Claude Code and Codex. The third is model routing, so you're not paying Opus prices for triage or simple CI fixes. You can use Warp's auto models or set your own rules, and Warp's internal evals found that UI tasks run well on GLM. Related links: Warp: https://www.warp.dev Timestamps: 0:00 Intro 0:50 Self-improving software factories 1:50 Three ways factories improve 2:05 Skills and outer-loop agents 3:15 Demo: Warp's triage skill 4:15 The outer-loop agent 5:00 Skill updates go through pull requests 5:35 Persistent memory 6:35 Demo: a Sentry agent's memories 8:15 Memory across harnesses 8:45 Model routing 9:40 Auto models 9:50 Demo: custom routing rules 10:50 Next: routing evals 11:40 How routing works inside Warp 12:25 Wrap-up