Understand what it tests
Behaviors define what the agent must get right. Tasks make those expectations concrete and replayable.
The open library for AI benchmarks
Discover, run, and improve AI benchmarks.
Benchmax organizes benchmarks into clear behaviors and tasks, so anyone on your team can understand what is being tested, run it against your agent, and improve it together.
Readable by the whole teamRuns with your harnessVersioned like code
Explore benchmarks
Evaluations for agents doing real work.
See the capability each benchmark measures before opening its behaviors, tasks, and published runs.
τ-Voice
Sierra Research
Full-duplex voice-agent evaluation across airline, retail, and telecom. Tests: Interruptions · turn-taking · policy adherence · tool use
Mar 2026
Mortgage Origination
Benchmax
Income and asset verification for mortgage agents reviewing borrower files. Tests: Income · assets · debts · source reconciliation
Aug 2026
Clinical Documentation
Benchmax
Evidence-grounded clinical notes generated from a complete visit transcript. Tests: Grounding · diagnosis support · medication accuracy
Aug 2026
Property Underwriting
Benchmax
Commercial property underwriting from submission intake to the appetite call. Tests: Source quality · conflicting facts · risk completeness
Aug 2026
How Benchmax works
A shared path from benchmark to better agent.
Run it against your agent
Connect your harness, run the tasks, and preserve results so the team can inspect what passed and why something failed.
Improve it together
Engineers, product managers, and domain experts can propose missing cases, review changes, and strengthen coverage.
Start with a benchmark.
Explore an open repository, understand what it tests, and see how Benchmax makes evaluation useful to the whole team.
Explore benchmarks