Agent-driven performance optimization of software systems
Performance work is detective work: find a clue, form a hypothesis, test it, follow the next clue when it fails. That is the same loop as agentic search, and much of the context sits in the project’s issues and PRs, not only in the code. Agents now do this well, from finding the fix to writing SIMD kernels and concurrent caches, on projects like my voice library mulive and an S3-backed vector database.
Takeaways
- Before fixing anything, mine the issues and PRs: maintainers may already know the problem, have half a fix, or have rejected the idea.
- Agents are strong at the code-level and algorithm-level work once they have the context.
- The hard part is the human side. A change gets accepted only when it fits the maintainers' mental model. Otherwise it gets called AI slop. That framing is the steering human's job.
Have a system that needs a second opinion?