Ghost Browser: Enabling Autonomous Agents to Click "Post"

Описание и применение

Ghost Browser позволяет автономным агентам выполнять реальные действия в браузере, обходя проблемы селекторов и управления сессиями, и обеспечивает подтверждение пользователя перед публикацией.

Задача

LLMs can plan actions but lack the ability to interact with real web pages, especially logged‑in actions like posting or messaging, due to brittle selectors and credential handling.

Результат

Agents can now perform real web actions reliably, surviving site redesigns, maintaining session persistence, and ensuring user approval for sensitive actions.

Рабочий процесс

• 1. Sign into target site manually once in Ghost Browser console. 2. Ghost Browser creates a persistent profile with cookies and storage. 3. Agent plans an action (e.g., "post a comment"). 4. Ghost Browser renders the page, overlays numbered boxes on interactive elements. 5. Agent receives annotated screenshot and list, then issues command like "click 12". 6. Ghost Browser executes the click in the live session. 7. For sensitive actions, act‑gate prompts user to approve, edit, or reject before execution.

Доказательства

Ghost Browser — a real Chromium an agent operates the way a person does.

Источник

The line that matters: it asks before it acts

Источник