Gemini APIのマネージドエージェントが拡張：バックグラウンドタスク、リモートMCPなどに対応 Expanding Managed Agents in Gemini API: background tasks, remote MCP and more

Google has expanded the managed agent capabilities within its Gemini API, adding support for background task execution and connectivity to remote Model Context Protocol (MCP) servers, among other features. The update matters because it shifts more of the operational complexity of building autonomous AI systems from the developer to Google's infrastructure, potentially lowering the barrier to producing agents that can plan, act, and manage state over longer time horizons.

Managed agents refer to a server-side abstraction in which the platform handles the orchestration loop that an AI agent depends on: interpreting a goal, calling tools, evaluating results, and deciding on the next step. Rather than requiring developers to write and host this loop themselves, the Gemini API manages the runtime, tool invocation, and conversational or task state. The newly announced additions extend what those managed agents can do and how long they can run.

The headline change is asynchronous background task execution. Historically, many API-based agent interactions have been synchronous, meaning a client sends a request and waits for a response within a single connection. That model works for quick question-and-answer exchanges but becomes fragile for workflows that take minutes or longer, such as researching across many sources, processing large datasets, or coordinating multiple tool calls. With background execution, an agent can appear to continue working after the initial request returns, allowing developers to kick off a job, poll for status, or receive results later. This pattern is a common prerequisite for building agents that behave more like durable workers than momentary chat responses.

The second major addition is remote MCP server connectivity. The Model Context Protocol is an open standard, originally introduced by Anthropic in late 2024, that defines how AI applications connect to external tools, data sources, and services through a consistent interface. It has since gained broad adoption across the industry, with support appearing in offerings from multiple major model providers and developer tool vendors. By allowing Gemini's managed agents to connect to remote MCP servers, Google is enabling agents to reach tools and data that live outside the immediate application, using a standardized protocol rather than bespoke integrations. In practice, this means a developer could point an agent at an existing MCP server exposing, for example, a database, a ticketing system, or an internal API, and have the agent use those capabilities without writing custom glue code for each one.

Taken together, these features address two persistent challenges in agent development: longevity and interoperability. Background tasks tackle the problem of workflows that outlast a single request, while remote MCP support tackles the fragmentation that arises when every integration must be built from scratch. Google frames the combination as giving developers stronger building blocks for constructing capable autonomous workflows, though the practical reliability of any given agent will still depend heavily on prompt design, tool quality, and guardrails.

The move fits into a broader industry pattern. Competing platforms have introduced comparable primitives, including OpenAI's Assistants and Responses APIs and its Agents framework, as well as agent-oriented tooling from Anthropic, Microsoft, and various open-source projects such as LangChain and LlamaIndex. The convergence around MCP as a connective layer is notable, because it suggests the market is coalescing on shared standards for how models talk to external systems, which could reduce lock-in and make it easier to move agent logic between providers. Google's decision to support remote MCP servers rather than only proprietary tool definitions is consistent with that trend.

There are important caveats. Autonomous agents that run in the background and connect to external systems raise questions around cost control, observability, security, and the risk of unintended actions, since an agent acting on its own can consume resources or invoke tools without a human reviewing each step. Developers adopting these features will likely need to consider authentication for remote MCP connections, monitoring for long-running tasks, and limits on what agents are permitted to do. As with earlier agent tooling, the gap between a compelling demonstration and a production-ready system remains significant, and organizations are likely to proceed incrementally. Developers interested in the specifics should consult Google's official Gemini API documentation for supported regions, pricing, and current limitations, which tend to evolve as such features move from preview toward general availability.