How should we benchmark Lightpanda for AI agents?

We ran Lightpanda, agent-browser, and browser-use through AssistantBench and GAIA Level 1 with Claude Sonnet 4.6 as the brain. The tool surface mattered more than the engine.