Benchmark·Coding agents
SWE-bench
SWE-bench
A benchmark for resolving real GitHub issues by generating patches that are checked against repository tests.
Benchmarks
Public AI Agent Benchmarks
Public task sets and environments that test whether agents can complete work through tools, code, browsers, or multi-turn interaction.
Showing 1-12 of 16 resourcesPage 1 of 2
Open resource
Benchmark·Browser agents
WebArena-x
WebArena
A realistic web environment for testing autonomous agents on web-based tasks across common site categories.
Benchmark·Computer-use agents
XLang Lab
OSWorld
A benchmark family for multimodal agents completing open-ended computer tasks across web and desktop applications.
Benchmark·Tool-agent-user interaction
Sierra Research
tau-bench
A benchmark for agents that converse with users, call tools, retrieve knowledge, and follow policies in enterprise-style domains.
Benchmark·Real-world work
OpenAI
GDPval
An evaluation for economically valuable professional tasks across 44 occupations.
Benchmark·Long-horizon work
Mercor
APEX-Agents
A benchmark for long-horizon, cross-application tasks in professional services environments.
Benchmark·Workflow orchestration
Zapier
AutomationBench
A benchmark for end-to-end workflow execution across realistic business tools and rules.
Benchmark·Legal agents
Harvey
Harvey LAB
An open-source benchmark for evaluating legal work in realistic document-and-tool settings.
Benchmark·Reasoning and knowledge
Center for AI Safety
Humanity's Last Exam
A hard benchmark for factual recall and knowledge calibration across broad domains.
Benchmark·Instruction following
Allen Institute for AI
IFBench
An instruction-following benchmark for checking whether models follow task rules.
Benchmark·Coding
LiveCodeBench
LiveCodeBench
A contamination-free coding benchmark that continuously collects new problems.
Benchmark·Science reasoning
GPQA
GPQA Diamond
A hard science reasoning benchmark used in frontier model comparisons.
Resource does not match your workflow?