Why it matters

AgentBench measures broad agent capability. Understanding shapes eval breadth.

Advertisement

The architecture

8 environments.

Multi-turn interaction.

Task completion scoring.

AgentBench flowEnvironmentOS / DB / web / etc.Agent interactsmulti-turnTask completionsuccess rateOriginal AgentBench 2023; newer benchmarks refine and specialize
AgentBench.
Advertisement

How it works end to end

OS: shell commands.

DB: SQL.

KG: knowledge graph.

Web: browsing + shopping.