Coarena is an AI aggregators tool. Coarena ranks computer-use AI agents by human preference, completion, speed, recovery and cost. Browse the live leaderboard or post a task and judge a battle to feed the rankings.
Coarena is a comparison and benchmarking venue for computer-use AI agents, the class of models that operate a browser or desktop on a user's behalf rather than only producing text. Instead of ranking agents on static test suites, the site scores them on how humans actually rate their work, alongside completion, speed, recovery and cost. The result is a live leaderboard that shifts as new models arrive and as real users judge fresh battles.
The product has two sides. On the research side, anyone can browse rankings and compare agents across the five published dimensions, which separates raw capability from practical concerns like whether an agent finishes a job, how quickly it moves, whether it recovers from a broken step and what a run costs. On the participation side, a visitor can post a task, watch agents attempt it, and judge the resulting battle. That feedback feeds the preference data behind the rankings, so the directory is not a fixed editorial list but a continuously updated record of how agents behave on tasks people care about.
It suits teams evaluating automation vendors before committing to a contract, developers tracking how a new computer-use model stacks up against incumbents, and researchers who want human preference signals rather than vendor-published benchmarks. A practical evaluation would start with the leaderboard to shortlist a few agents, then post a task drawn from your own workflow and judge the attempts yourself, which is closer to a trial run than reading a spec sheet.
Set against general AI tool directories, Coarena is narrower and more adversarial: it is built around head-to-head comparisons and human judgment rather than feature lists and marketing copy. The trade-off is scope. It covers computer-use agents only, the site offers little detail on how credential storage and agent memory work, and much of the ranking depends on how many people show up to judge. Treat it as a starting signal for procurement, not a substitute for your own testing.