The Streamlit team shares how they redesigned their 8-year-old open-source Python codebase to work effectively with AI coding agents. Rather than relying on better prompts or models, they built a layered 'harness' system: simplifying the codebase to reduce agent confusion, tightening linting/type-checking guardrails across Python/TypeScript/Protobuf, moving all project knowledge (AGENTS.md files, architecture docs, specs) into the repo, packaging recurring work into reusable skills and subagents, adding multi-model AI PR review, and automating maintenance tasks via GitHub Actions. The result: features that took a sprint now ship in hours, newly reported bugs are at a three-year low, and agents handle the majority of feature work and maintenance while humans focus on product, architecture, and design decisions.
