Probe to Act lets browser agents inspect uncertain targets before taking state-changing actions. Agents can render DOM candidates onto screenshots, map visual regions back to DOM elements, register visual-only targets, and save verified observations.
The framework works through prompting for proprietary models and through synthesis plus self-bootstrapped supervised fine-tuning for open-weight models. On VisualWebArena, success rose from 54.1% to 61.2% for Gemini-3-Pro and from 24.6% to 32.9% for Qwen3-VL-8B.
Its evidence-based memory retains only probed, acted-on, or explicitly committed information. It reportedly matches full-observation history while using about 1.2 times the peak retained input context of action-only history, rather than about 3 times for full history.