Evaluation & metrics

What benchmarks exist for evaluating AI agents?

Short answer

There are several public benchmarks for agents, and each measures a different thing. SWE-bench tests coding agents on real GitHub issues. WebArena tests agents that operate a browser. tau-bench tests tool use and conversation in realistic customer-service settings. GAIA tests general assistants on multi-step reasoning. AgentBench and the Berkeley Function-Calling Leaderboard test broad agent and tool-calling ability. They are useful for comparison, but your own scenario set still matters more for your specific use case.

The benchmarks worth knowing

  • SWE-bench measures whether a coding agent can resolve real GitHub issues with a working patch.
  • WebArena tests agents that navigate and act inside realistic websites.
  • tau-bench evaluates tool use and multi-turn dialogue in customer-service style tasks with shared state.
  • GAIA probes general assistants on multi-step, real-world reasoning questions.
  • AgentBench spans many environments to measure broad agent capability.
  • Berkeley Function-Calling Leaderboard ranks models specifically on function and tool-calling accuracy.
  • ToolBench focuses on tool-use ability across many real APIs.

What benchmarks are good for

Public benchmarks let you compare models and track the field. They are a sanity check on raw capability, and a fast way to rule a model in or out.

What they are not

A high benchmark score does not mean the agent works for your product. Benchmarks cannot know your data, your permissions, or your definition of success. Treat them as a starting filter, then rely on your own golden dataset and workflow tests for decisions that matter.