Coarena by Coasty
Coarena operates a live arena where real people post browser tasks and two computer-use agents race the identical task in parallel while a human judge evaluates the outcome. The platform enforces blind judging by scrubbing vendor names at the database write layer, and every published metric cites the specific source file and rule that produced it. The arena produces trajectory datasets, leaderboard rankings, and an API for accessing metrics.
- Artificial Intelligence
- Data & Analytics
- AI Agents
- Developer Tools
- Software Only
Funding
Raised to date
$500KRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Founder details are not available yet.
Product
Problem
Fixed computer-use benchmarks degrade over time: tasks leak into training corpora, scores crowd the ceiling, and rankings freeze while models continue improving. Published benchmarks become targets that agents overfit, making their numbers unreliable for comparing modern computer-use agents.
Solution
Coarena operates a live arena where real people post real browser tasks and two computer-use agents race the identical task in parallel while a human judge evaluates the outcome. The platform enforces blind judging by scrubbing vendor names at the database write layer, and every published metric cites the specific source file and rule that produced it. Tasks are screened for overlap with public benchmarks to prevent leakage, and all agents compete on a shared action vocabulary, browser, sandbox, step budget, and wall clock. The arena produces trajectory datasets, leaderboard rankings, and an API for accessing metrics, with safety benchmarks and model comparison tools included.
Target Audience
Coarena serves AI labs and model developers who need reliable comparative evaluations of computer-use agents, as well as model buyers and researchers seeking inspectable, leakage-resistant performance data.
Features
- Live task submission where real users post browser tasks that two agents race in parallel under identical conditions
- Blind judging enforced at database write time, with two scrubbers for code-authored and model-authored text, plus fail-closed identity visibility logic
- Task screening pipeline that checks submissions against public computer-use benchmarks to prevent training-data leakage
- Shared action space with one action vocabulary, one browser, one sandbox, one step budget, and one wall clock per lane
- Trajectory dataset containing agent actions, reasoning, and page states, with personal data redaction and face-template detection
- Metrics API and published methodology covering 101 defined measures, with each claim citing the source file that enforces it
- Public result cards for individual battles showing task prompt, model names, final answers, and winner only