How Software Factories Improve Themselves — Suraj Gupta, Warp

This post details three practical methods for self-improvement within software factories, exemplified by Warp's open-source factory. First, 'skills that improve themselves' are achieved via an outer-loop agent that monitors triage agent runs and feedback, then opens PRs to update skills, ensuring Git tracking and human review. Second, 'persistent memory' provides a versioned, traceable store of facts, preventing agents (like a Sentry agent) from re-discovering known root causes. This memory system is harness-agnostic. Third, 'model routing' optimizes costs by directing tasks to appropriate models (e.g., using cheaper auto models for triage or simple CI fixes instead of Opus), with Warp's internal evals showing UI tasks perform well on GLM.