PerfAgent triples optimization success by reading profiler traces

Coding agents crush correctness benchmarks and then fall apart on repository-level performance work. The failure mode is specific: an agent finds a patch that passes tests and stops, even when a profiler would show the real bottleneck sitting one layer down, hidden behind an abstraction it never inspected.

PerfAgent fixes this by wrapping an existing agent in a profiler-guided loop with a verifier in the middle. Instead of picking the next optimization attempt based on wall-clock timing, it picks based on what the profiler actually flags. Running OpenHands with GPT-5.1 through this loop took GSO scores from 19.6% to 39.2%, and SWE-fficiency-Lite from 26% to 74%, per the paper (arXiv 2607.19653).

The part that should bother anyone benchmarking agents is the cost comparison. PerfAgent beat an oracle best-of-five baseline, the setup where you sample five attempts and hand-pick the best one, at lower cost. That's not a scale story. Best-of-five is brute force: burn more compute, take more shots, hope one lands. PerfAgent won by improving the signal each attempt is optimized against, not by taking more attempts.

That distinction matters for how builders spend their compute budget. If your agent workflow is stuck on performance regressions or repository-level optimization, the instinct is to widen the sampling, run the same prompt five or ten times and pick a winner. This result says the cheaper fix is upstream: give the agent profiler evidence as its feedback signal instead of pass/fail, and it stops settling for the first patch that clears the bar.

The paper doesn't say whether this generalizes past GSO and SWE-fficiency-Lite, or whether the gains hold on languages and runtimes without mature profiling tooling. Worth watching before assuming it ports to your stack.