Experiment: DOMShell vs Vision (browser-use) with same local model

[apireno]

Summary

Run the same Wikipedia tasks from experiments/nexa/ using Nexa's vision-based Web-Agent-Qwen3VL (Playwright + screenshots) and compare against the DOMShell results.

Why

The current Nexa experiment compares 1.7B/4B local models against Claude Opus — that's a model-size comparison, not an interface comparison. To validate DOMShell's text/AX-tree design, we need an apples-to-apples test: same model, same tasks, different interface (DOMShell text vs Playwright screenshots).

Approach

• Use Nexa's existing Web-Agent cookbook (Playwright + browser-use + Qwen3-VL)
• Run the same 3 Wikipedia tasks from experiments/nexa/nexa_prompts.md
• Compare: tool calls, correctness, completeness, token usage
• Key question: does DOMShell's structured text approach outperform pixel-based browsing at the same model size?

Expected Outcome

DOMShell should be more token-efficient (text vs images) and may enable smaller models to succeed where vision models need 4B+ parameters.

References

• experiments/nexa/ — existing DOMShell results (0/12 with 1.7B-4B)
• integrations/nexa/agent.py — DOMShell agent
• Web-Agent-Qwen3VL cookbook — vision-based comparison target