How do we measure agent capabilities?

How do we measure agent capabilities?

Instead of pairwise preference voting, Agent Arena uses a causal tracing methodology that measures five signals from real-world, long-horizon agentic sessions.

More info in thread.