Multi-domain benchmarking: AgentBench-Live covers 5 domains beyond coding

[jackjin1997]

PinchBench looks great for evaluating coding agents specifically.

We built AgentBench-Live which extends the concept to 5 domains: Code, Data Analysis, Research, Tool Use, and Multi-step Workflows. Each domain has its own scoring methodology (auto tests, LLM-as-Judge, composite).

Key insight from our benchmarks: code tasks are essentially "solved" (both Claude Code and Gemini CLI score 1.00), but agents diverge significantly on complex tasks requiring planning, data interpretation, and tool orchestration.

Would be great to explore integrating PinchBench's coding benchmarks as a domain in AgentBench-Live, or vice versa. Open to collaboration!

[olearycrew]

I think a collaboration could be cool - but to be clear PinchBench is intentionally focused on domains that go outside of coding alone to try and represent more real-world use cases for OpenClaw. So not sure if that was an automated message that said

integrating PinchBench's coding benchmarks as a domain in AgentBench-Live

But interested in what specifically you're thinking a collab would look like in terms of both technical implementation and what it would look like to users of either of our benchmarks