Built an AI orchestration system that writes PRDs, ships code and reviews PRs. Along the way I replaced myself. I added one agent, then another. The factory must grow. #buildinpublic#AI
Part IV: The Factory Must Grow. Building an AI orchestrator that ships PRs without me. Added End-to-End (E2E) testing that show agents are doing what they're supposed to #buildinpublic#AI
Part III: The Factory Must Grow. My AI agent factory silently dropped dozen PRDs and auto-closed them. The fix wasn't a plaster it was learning to stop the line completely and walk every failure to a policy. Dead code beats limping code. A work in progress...! #buildinpublic#AI
My AI agent orchestrator grew into spaghetti: silent fallbacks, dropped issues, PR ping-pong burning Claude tokens. The fix came from Elm: a pure reducer that decides, an interpreter that does. Dropped issues are impossible by design. #buildinpublic#AI
4/ Impossible by design.
The reducer handles every event type. The TypeScript compiler refuses to build if a case is missed. The "missing branch" bug that silently dropped issues for weeks can't be written anymore.