PI AutoResearch automates performance tuning loops

This is less a flashy AI demo than a practical experiment harness, and that’s what makes it interesting: it turns optimization into a disciplined loop with commits, measurements, and rollback baked in.

• –The split between global extension and per-domain skill is the right architecture: reusable mechanics stay generic while the optimization goal stays customizable.
• –The session files and JSONL history make the workflow reproducible, so the loop can survive restarts and context resets without losing state.
• –It’s broader than “AI coding” in the usual sense; the same pattern can chase regressions in tests, builds, model training, or web perf.
• –The biggest risk is metric quality: if the benchmark or backpressure checks are weak, the agent can optimize the wrong thing very efficiently.