Towards Self-Driving Codebases
4 hours ago
- Agents can perform fun one-shot games and complex migrations, but most real software work still requires human engineers.
- Current attempts to offload work to agents have produced disappointing results, leading to a trough of disillusionment.
- Future engineering value will come from having good ideas (features, architecture) rather than setting up agent loops.
- Self-driving areas should include bug detection and fixing, debugging production errors, optimizing agents, frontend consistency, application polish, and growth optimization.
- Missing primitives include agent-legible dev environments, global memory to avoid repeated mistakes, and codebase rot prevention.
- Engineering teams should invest in making dev environments agent-friendly by mining codebases for bugs and fixing gaps.
- The path forward involves using bug traces to prioritize investments, benchmarking agent readiness, and progressively handing off work.
- This mirrors historical technology shifts like cloud migration, where missing primitives were eventually solved.