PerfAgent: Profiler Feedback Doubles Expert-Matching Optimization Patches — 26% to 74% on SWE-fficiency-Lite

arXiv 2607.19653·high signal

Coding agents are good at correctness benchmarks like SWE-bench but weak at repository-level performance work, where they miss bottlenecks behind abstraction layers and stop at the first passing patch. PerfAgent wraps an off-the-shelf agent with a profiler-guided, verifier-in-the-loop workflow that picks the next optimization from profiler evidence rather than wall-clock timing, lifting expert-matching patches from 19.6% to 39.2% on GSO and 26% to 74% on SWE-fficiency-Lite over OpenHands with GPT-5.1. It beats an oracle best-of-five baseline at substantially lower cost, meaning the win comes from better feedback signal, not more sampling.