Benchmark
AgentBench
Evaluate language-model agents reasoning and acting across interactive environments.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
The original v0.2 suite covers eight environments: operating systems, databases, knowledge graphs, games, puzzles, household tasks, web shopping, and web browsing.
Methodology
Multi-turn interactions assess reasoning and decisions. Three environments adapt ALFWorld, WebShop, and Mind2Web. The current repository defaults to AgentBench FC, a separate function-calling setup with five containerized environments; keep its results distinct from the original suite.