FrontierHarness Eval

FrontierHarness Eval provides comprehensive evaluation of 9 harnesses across 12 configurations, tested on identical software engineering tasks using the same model and runtime. It addresses the challenge of comparing AI coding agent performance, pass rates, costs, and runtimes without warm-cache bias or home-field advantage.

Designed for developers, researchers, and teams building or deploying AI coding harnesses, the tool guarantees that all 360 trials start from the same fresh checkpoint restore. It provides clear metric breakdowns including quality, speed, median cost per task, and cache hit rates to help users understand the real tradeoffs between different agent execution layers.