Google Deepmind adds background execution and MCP support to Gemini API managed agents Google DeepMind 为 Gemini API 托管代理添加后台执行和 MCP 支持

TL;DR

• Google DeepMind introduces Background Execution for Gemini API Managed Agents, enabling asynchronous processing without maintaining open HTTP connections.
• Support for remote Model Context Protocol (MCP) servers allows direct connectivity to internal databases and APIs.
• Developers can now integrate custom functions alongside built-in sandbox tools within the agent environment.
• Credential refresh capabilities enable token updates between interactions while preserving the sandbox state.

Why It Matters

These updates significantly enhance the scalability and flexibility of building autonomous AI agents by decoupling execution from immediate client connections and expanding integration capabilities with external systems. For developers, the ability to run agents asynchronously and manage credentials securely reduces infrastructure complexity and improves reliability in long-running tasks.

Technical Details

• Background Execution: Implements asynchronous agent processing via the Gemini Interactions API, eliminating the need for persistent HTTP connections during agent operation.
• MCP Integration: Facilitates direct connections to remote Model Context Protocol servers, allowing agents to interact with internal databases and third-party APIs seamlessly.
• Custom Functionality: Expands the tooling ecosystem by permitting the use of user-defined custom functions in conjunction with existing sandbox tools.
• Stateful Credential Management: Introduces mechanisms to refresh authentication tokens between interactions without resetting or losing the current sandbox state.
• Cross-Language Support: Documentation provides code examples for JavaScript, Python, and cURL to assist implementation across different development environments.

Industry Insight

The introduction of asynchronous execution and MCP support signals a shift toward more robust, enterprise-grade AI agent architectures that can handle complex, multi-step workflows without tight coupling to client sessions. Practitioners should prioritize integrating these features to improve agent reliability and reduce latency in production environments. Furthermore, the ability to maintain state during credential refreshes addresses a critical pain point in secure, long-lived agent deployments, encouraging broader adoption in regulated industries.

TL;DR

• Google DeepMind 为 Gemini API 托管代理引入了后台异步执行功能，无需保持开放的 HTTP 连接即可运行。
• 新增对远程 MCP（模型上下文协议）服务器的支持，允许代理直接连接内部数据库或 API。
• 开发者现在可以将自定义函数与内置沙箱工具结合使用，并支持在交互间刷新凭证而不丢失沙箱状态。
• 所有新功能均通过 Gemini Interactions API 提供，并附有 JavaScript、Python 和 cURL 的代码示例。

为什么值得看

此次更新显著提升了 Gemini 托管代理在企业级应用中的实用性和稳定性，特别是后台执行和 MCP 支持解决了长期运行的自动化任务痛点。对于希望将 AI 代理集成到现有后端系统或需要处理敏感数据流式的开发者而言，这些功能是构建生产就绪型应用的关键基础设施。

技术解析

• 后台异步执行 (Background Execution)：允许代理在无活跃 HTTP 连接的情况下在后台运行，降低了资源占用并提高了长时间任务的可靠性，适合处理耗时较长的复杂工作流。
• MCP 服务器集成：原生支持连接远程 Model Context Protocol 服务器，使得 Gemini 代理能够更灵活地访问外部数据源、内部数据库或第三方 API，增强了上下文获取能力。
• 自定义函数与沙箱共存：打破了仅依赖内置工具的局限，开发者可以在受控的沙箱环境中混合使用自定义代码逻辑和预定义工具，提升了功能的灵活性。
• 状态持久化的凭证刷新：解决了传统会话中凭证过期导致状态丢失的问题，支持在多次交互间安全刷新令牌，同时保持沙箱环境状态的连续性。

行业启示

• Agent 工程向生产化迈进：随着后台执行和状态管理的完善，AI 代理正从简单的对话演示转向可信赖的、长期运行的企业级自动化服务。
• MCP 成为标准接口：Google 对 MCP 的支持表明该协议正在成为连接 LLM 与外部数据/工具的事实标准，加速了异构系统间的互操作性。
• 开发体验优化：通过提供更接近传统后端开发模式（如异步调用、凭证管理）的功能，降低了非 AI 专家构建复杂智能系统的门槛。