Two AI Agents, One Browser: The Race Condition Behind Computer Use

Описание и применение

Многопользовательская автоматизация браузера с помощью AI приводит к гонкам ресурсов. Mac MCP использует стабильные идентификаторы вкладок, координацию совместного доступа и фоновые безопасные действия, чтобы уменьшить ошибки и улучшить надёжность.

Задача

Concurrent AI agents interacting with the same browser can cause race conditions, leading to incorrect actions, lost state, and user confusion.

Результат

By treating tab positions as non‑identifiers, coordinating shared resources, and distinguishing background actions, Mac MCP reduces race conditions and improves reliability of multi‑agent desktop automation.

Рабочий процесс

• 1. Agent opens a tab and performs an action using a stable tab handle.
• 2. A second agent may open, close, or reorder tabs, potentially changing the tab’s position.
• 3. Agents coordinate on shared resources (e.g., forms) to avoid concurrent writes.
• 4. Background‑safe tools target tabs without switching focus, preserving user attention.
• 5. After an action, the system reports three states: action attempted, outcome observed, outcome uncertain.
• 6. Tests deliberately trigger interruptions (tab reorder, close, concurrent writes) to validate robustness.

Доказательства

An array position was used as if it were a durable resource identifier.

Источник

Stable handles solve one problem, but they don't resolve concurrent writes.

Источник

Mac MCP's ordinary Safari and Chrome browser work is designed to target background tabs without switching the user's active tab.

Источник

I try to keep three different states separate: Action attempted, Outcome observed, Outcome uncertain.

Источник

The most useful tests for this kind of tool are deliberately awkward: Open, close, and reorder tabs during an agent action.

Источник