Google DeepMind has unveiled a comprehensive suite of updates for its Managed Agents within the Gemini API ecosystem, marking a significant milestone in the evolution of autonomous artificial intelligence. The announcement, spearheaded by Developer Relations Engineer Philipp Schmid and Product Manager Mariano Cocirio, introduces four primary capabilities designed to transform AI agents from simple conversational interfaces into robust, production-ready workers. These updates include long-running background execution for asynchronous interactions, direct integration with remote Model Context Protocol (MCP) servers, support for custom function calling alongside sandbox tools, and the ability to refresh network credentials across interactions.
By addressing critical developer feedback regarding the fragility of long-running HTTP connections and the complexity of connecting to private data silos, Google is positioning its Gemini Interactions API as a central hub for agentic workflows. With managed agents, developers can now call a single endpoint while Google handles the heavy lifting of reasoning, code execution, package installation, and file management within an isolated cloud sandbox. This release signals a shift toward "Agentic AI," where models are no longer just predicting text but are actively managing software development lifecycles and internal enterprise operations.
The Evolution of Agentic Infrastructure
The introduction of these features follows a period of rapid iteration within the Google AI DevRel team. Historically, AI models operated in a "stateless" manner, where each request was independent and required the client to manage the context and history. The launch of the Gemini Interactions API earlier this year began to change this by providing a managed environment. However, developers still faced hurdles when tasks took longer than a standard web timeout or when agents needed to access secure, internal databases.
The current update addresses the "Antigravity" preview models, specifically designed to function as autonomous agents that can clone repositories, analyze source code, and generate complex reports without constant human intervention. This progression aligns with broader industry trends observed throughout 2024, where the focus moved from the size of the Large Language Model (LLM) to the efficacy of the "agentic loop"—the process by which an AI plans, acts, observes, and corrects its own behavior.
Long-Running Background Execution and Asynchronous Stability
One of the most technically significant additions is the support for background execution. In previous iterations, developers often struggled with "fragile" HTTP connections. If an agent was tasked with a complex job—such as scanning a massive GitHub repository for security vulnerabilities—the connection would often time out before the agent could finish its reasoning and execution.
Under the new system, developers can pass a background: true parameter. The API immediately returns a unique interaction ID, allowing the client application to disconnect. The agent continues to work remotely in the Google-managed sandbox. Applications can then poll for status, stream progress updates, or reconnect hours later to retrieve the final output. This asynchronous model is essential for enterprise-grade applications where reliability is paramount. It mirrors the transition in traditional software engineering from synchronous API calls to message-queue architectures, which are better suited for heavy workloads.
Bridging Private Data Gaps with Remote MCP Integration
Perhaps the most impactful update for enterprise developers is the integration of the Model Context Protocol (MCP). Originally introduced as an open standard to allow AI models to interact with various data sources, MCP has quickly gained traction across the industry. Google’s implementation allows managed agents to connect directly to remote MCP servers.
Previously, accessing a private database or an internal telemetry server required developers to write custom proxy middleware to "feed" data to the AI. With this update, an agent can communicate with an internal endpoint from within its secure sandbox. This capability allows developers to mix and match tools: an agent could use Google Search to find external documentation, code execution to write a script, and a remote MCP server to query real-time latency spikes from an internal auth service.
This integration significantly reduces the "glue code" required to build functional agents. By providing a standardized way to access internal APIs, Google is lowering the barrier to entry for companies that want to deploy AI agents on top of their proprietary data stacks without compromising security.
Hybrid Tooling and Custom Function Calling
The third pillar of the update is the enhancement of custom function calling. Google has introduced a "step matching" system that allows built-in sandbox tools to run alongside local business logic. In this hybrid model, tasks that the sandbox can handle—like running Python code or managing files—are executed automatically on Google’s servers. However, when the agent identifies a need for a custom domain function (such as checking a specific local weather API or querying a local ERP system), the interaction transitions to a requires_action state.
This state-based approach allows for a "human-in-the-loop" or "local-system-in-the-loop" workflow. The client-side application executes the local tool and sends the result back to the agent to continue its reasoning. This ensures that while the AI has the power of a cloud sandbox, it remains tethered to the specific, authorized functions of the host application.
Security Persistence through Credential Refresh
As agents move into production, security and session management become critical. Many modern APIs use short-lived access tokens or OIDC (OpenID Connect) tokens that expire within minutes or hours. In a long-running agentic session—where an agent might be working on a project for several days—token expiration used to be a session-ending event.
Google’s new "network credential refresh" capability allows developers to rotate keys or refresh tokens without destroying the agent’s environment. By passing an existing environment_id with new network configuration rules, the agent can continue its work seamlessly. The underlying sandbox keeps its filesystem state, installed packages, and cloned repositories intact, but gains the new authorization headers required to continue accessing external resources like Google Cloud Storage (GCS) or private GitHub repos.
Market Analysis and Industry Implications
The move by Google DeepMind to enhance managed agents is a direct response to intensifying competition in the AI developer space. Competitors such as OpenAI, with its Assistants API, and Anthropic, with its tool-use capabilities, are all vying to become the primary platform for the next generation of AI-driven software.
According to recent industry data from Gartner, by 2026, at least 15% of daily work decisions will be supported by autonomous agents. Google’s focus on the "managed" aspect—providing the sandbox, the execution environment, and the networking rules—is a strategic attempt to capture the enterprise market. Enterprises are often hesitant to allow AI models to execute code locally on their own infrastructure due to security risks. By providing an isolated cloud sandbox, Google offers a "safety-first" alternative that satisfies IT security requirements while giving developers the flexibility they need.
Industry analysts suggest that the integration of MCP is a particularly savvy move. "By adopting open protocols like MCP, Google is acknowledging that the future of AI is not a walled garden," noted one independent tech analyst. "The value isn’t just in the model; it’s in how easily that model can talk to the rest of the world’s data."
Timeline of Gemini API Development
The roadmap leading to this announcement shows a clear trajectory toward autonomy:
- December 2023: Introduction of Gemini 1.0 and the initial Gemini API.
- February 2024: Launch of Gemini 1.5 Pro with its massive 1-million-token context window, enabling long-context reasoning.
- May 2024: Introduction of the Gemini Interactions API and the concept of managed sandboxes.
- Late 2024: Beta testing of "Antigravity" agent models designed for complex coding tasks.
- Present: Expansion into asynchronous background tasks and standardized remote data protocols.
Conclusion and Forward Outlook
The updates to Managed Agents in the Gemini API represent a maturation of AI technology. We are moving away from the "chatbot" era and into the "worker" era. For developers, these tools mean less time spent managing infrastructure and more time spent designing the logic and goals of their agents.
As these capabilities move from preview to general availability, the next challenge for Google and its peers will be the refinement of "agentic reliability"—ensuring that as agents become more autonomous, they remain predictable and cost-effective. For now, the ability to run tasks in the background and connect to private servers via MCP provides the essential plumbing needed to build the autonomous AI applications of tomorrow.
Google has encouraged developers to begin experimenting with these new features via the @google/genai JavaScript SDK or the Python equivalent. With the ability to maintain stateful environments and refresh security credentials on the fly, the path is now clear for agents to take on more significant roles in software engineering, data analysis, and automated research.
