Which harness is the best coding agent?
An agent = a model + a harness. The harness is the scaffolding — the CLI or TUI — that wraps a raw model and turns it into something that edits files, runs tools, and finishes a task. Top Model holds the harness fixed and ranks models; here we flip it and rank the harnesses themselves — each scored by the best results it gets across the models we ran on it.
Harness leaderboard
A single-model agent is still a real leaderboard entry — the Models column surfaces how many models we ran on each harness, so you can see coverage at a glance. Score is the best composite a harness reaches, so a harness is only as good as the models it can drive.
Known agents · not yet benchmarked
8 · popular on OpenRouterWidely-used coding agents that haven't submitted to a benchmark we scrape (SWE-bench, Terminal-Bench, CodingAgentBench), so they have no harness × model scores yet — they can't be ranked until they do.