High-Speed, Zero-Token-Waste Autonomous Browser Engine for AI Agents

Описание и применение

Модуль agent-browser-mesh снижает токенов и время выполнения автономных браузерных агентов, заменяя снимки экрана и большие DOM‑дампы семантически сжатой передачей изменённых узлов, при этом обеспечивает защиту от внедрения запросов.

Задача

Full-page screenshots for vision LLMs consume ~1,500 tokens per action and 3–6 s per step, creating a Vision Token Tax that hinders multi-step web scraping, form filling, and research tasks.

Результат

Agent-browser-mesh reduces token consumption by 92 % and step latency by 95 % compared to screenshots or raw HTML dumps, while adding robust prompt‑injection protection.

Рабочий процесс

• Install agent-browser-mesh via pip
• Instantiate BrowserMesh with BrowserConfig(headless=True, block_trackers=True)
• Navigate to target URL with mesh.navigate
• Retrieve semantically grounded interactive elements via page.get_interactive_elements()
• Execute actions (e.g., click_element) with safety inspection

Доказательства

When building autonomous browser agents with Playwright, Selenium, or Puppeteer, taking full-page screenshots and feeding them to vision LLMs burns over 1,500 tokens per action and takes 3 to 6 seconds per step.

Источник

Only transmits DOM nodes that actually changed between agent steps, cutting token consumption by over 92%.

Источник

Sub-Millisecond Security Gate (<0.4ms): Intercepts indirect prompt injection payloads hidden inside HTML comments, hidden divs, and image alt tags before they enter the LLM reasoning loop.

Источник

Token Cost per Step: ~1,600 tokens (screenshots) vs ~180 tokens (agent-browser-mesh) (-92%)

Источник

Step Latency: 3,200ms (screenshots) vs 140ms (agent-browser-mesh) (-95%)

Источник