Test the work that matters
Choose a registered HUD or Harbor benchmark. Run different agent scaffolds in isolated environments, with a grader for every task.
Agent evaluation, with the evidence.
Know how your agent really performs.
Run the tests. Understand the failures. Build a better agent.
Benchmark agents on the same tasks and turn their results into private reports—with the detail your engineers need and the clarity your customers expect.
Private by default Evidence for every result
BLOBFISH / REPORTSSAMPLE DATA
Browser workflows · 12 example tasks
Competitive landscape
3agent configurations4task categories
AgentTask successCost / task
Atlas83.3%$0.101
Your agentYOU66.7%$0.071
Orbit50.0%$0.046
Your next improvement is in the details.Open any failed task to see the output, evaluator explanation, and execution trace.
Task success & reliability Cost & latency Failure analysis Raw outputs, grader verdicts & reports
From evaluation to improvement
A score is the start.
The report tells you what’s next.
Find the failure patterns
Separate incorrect outcomes from timeouts and execution errors. See the category breakdown, then inspect the exact task evidence.
Make the next release better
Compare configurations under matching conditions. Give engineering a focused starting point and share a report when you’re ready.
Two views. One body of evidence.
Reports built to be used.
01 / SINGLE AGENT
The performance report.
Go from an overall score to the exact task where your agent got stuck.
- Success, reliability, latency, and cost
- Category breakdown and failure patterns
- Searchable task outputs and evaluator explanations
66.7%Task success100%Reliability26.4sAvg latency
Open sample report 02 / COMPETITIVE LANDSCAPE
See where you stand.
Compare agents on matching task sets, with the context behind the ranking.
- Performance and cost comparison
- Category heatmap and agent-by-agent failures
- Clear methodology and coverage for each run
Atlas 83%Your agent 67%Orbit 50%
Open competitive sample One workflow, many benchmarks
Your benchmark.
Your scaffold. Clear results.
HUD FORMAT
HUD tasksets
Run portable environments with OpenAI or Claude agent scaffolds. Keep the native task grade and full trajectory.
Prepared environment images · pinned SDKExplore the workflowHARBOR FORMAT
Harbor benchmarks
Use task.toml, instructions, environment images, and tests/test.sh. Compare Terminus 2, Codex, and Claude Code.
A fresh task environment for every attemptExplore the workflowVISION + RUBRIC
Browser evaluation
Register browser tasks with screenshot evidence and task-specific rubrics. Export verdicts and reasoning for each task.
WebVoyager and Mind2Web require prepared task packagesExplore the workflowThe live registry shows exactly which task packages are available. Infrastructure controls are labeled separately from capability benchmarks. Every report retains its dataset version, task coverage, scaffold, and grading rules.
Your next release, backed by evidence.
Find the gap.
Then close it.
Start an evaluation or turn an existing run into your first report.