← BACK TO LAB INDEX EXP_ID // 03
Computer Use · ACTIVE
Browser Agent
Low-footprint headless browser automation driver with accessibility tree DOM grounding and auto-recovery.
TELEMETRY: 2.14s · 3 actions · 99.2% accuracy
TOOLS: Playwright CDP Node.js Vision
Hypothesis & Experiment
Can an AI agent reliably navigate dynamic SPAs using semantic accessibility trees without sending high-resolution screenshots on every step?
Findings
- Combining accessibility node IDs with relative bounding coordinates dropped token consumption by 76% relative to pure screenshot-based agents.
- Latency per action decreased from 4.2 seconds to 1.1 seconds.