Autonoma co-founder Tom Piaggio has published a candid post-mortem explaining why the AI-powered QA testing startup decided to throw out 18 months of production code — code that was actively serving paying customers — and start from scratch.
Two things broke at once. The first was self-inflicted technical debt. Piaggio ran a "no strict, no tests" TypeScript monorepo from the beginning, a tradeoff that worked fine when two founders knew every line of code. It didn't survive the company growing to 14 engineers. Null errors and undefined behaviour became routine, and Piaggio says the instability ultimately cost them at least one client. The rewrite goes in with strict TypeScript and test coverage as baseline requirements.
The second problem is more revealing about where the AI tooling industry sits right now. Autonoma's original architecture was built around the limitations of GPT-4 before the 4o release — a model that needed considerable hand-holding to reason about UI state. The team built Playwright and Appium inspection wrappers with seven self-healing click strategies and deep DOM introspection pipelines to compensate for what the model couldn't do on its own. By the time modern frontier models shipped, that scaffolding had become a liability. The models no longer needed the workarounds, and maintaining the old wrappers was blocking the fully agentic architecture Piaggio says they'd wanted to build from day one.
The infrastructure stack is changing too. Next.js Server Actions are out — Piaggio takes issue with sequential global execution, poor testability, and what he describes as security footguns — replaced with React, tRPC, and Hono. For workflow orchestration the team evaluated Temporal and useworkflow.dev before landing on Argo, primarily for its Kubernetes-native design.
Autonoma's situation is a useful case study in a problem that's going to keep showing up. Companies that built products in 2023 and early 2024 were, by necessity, building around the models available then. Some of that engineering was genuinely clever. A lot of it is now just in the way. The teams that can recognise when their scaffold has stopped being an asset are the ones that get to build the next version of the product.