Build Your Own Team of Managed Agents - Jack Wotherspoon, Google DeepMind

Highlights

• Cloud or managed agents run in cloud sandboxes rather than on a local machine, eliminating issues with browser hijacking and resource contention.

• The Interactions API and environments serve as the two core stateful backend primitives for managed agents, allowing separated conversation states and file system reuse.

• Gemini 4 introduced a long decode capability sharing a 1 million context window across input, thinking, and output for intensive math and financial modeling.

• Secrets never enter the sandbox directly because an API proxy acts as an intermediary, injecting placeholder tokens and matching them to requests securely.

• Managed agents support scheduled triggers for automated tasks like morning financial analysis, as well as multi-agent workflows where separate agents build on shared file environments.

Timeline

1. 00:00 - 02:41

   Introduction to Managed Agents

   Running coding agents locally presents friction because verification processes often take over the browser and trigger unwanted pop-ups. Personal laptop configurations and customized agent skills also lack the scalability required for team collaboration or customer-facing applications. Transitioning these operations to cloud sandboxes resolves these infrastructure constraints.

   • Cloud or managed agents execute coding tasks within cloud sandboxes instead of local developer machines.
   • Sandboxes prevent local issues such as browser takeover and pop-up interference during automated verification.
   • Running custom coding setups locally does not scale well when sharing with colleagues or deploying into customer applications.
2. 02:41 - 05:22

   The Interactions API and Architecture

   Building managed infrastructure manually requires provisioning containers, managing network security, and writing custom harnesses. Google DeepMind packages these components into a single API call within the Interactions API. This client library mirrors multi-step agent behavior, allowing models to write files, invoke tools, and leverage server-side capabilities like Google Search.

   • Managed agents are invoked through a single API call that bundles the sandbox and the anti-gravity harness out of the box.
   • The Interactions API replaced traditional single-turn model inputs with an agent-first client library supporting multi-step tool calls.
   • Under the hood, the Gemini API utilizes server-side search and file system tools alongside Google Cloud infrastructure sandboxes.
3. 05:22 - 08:50

   Core Pillars: Harnesses, Skills, and Credentials

   Managed agents rely on several foundational pillars, including the anti-gravity harness used across Google product suites. Model selection allows tuning for speed using Flash or deep reasoning using frontier models like Gemini 4 Argon. Furthermore, automated triggers allow agents to run on a set schedule every morning to process emails, scan stock markets, or query project boards.

   • The core pillars of managed agents consist of the anti-gravity harness, isolated environments, uploaded skills, credentials, and triggers.
   • Gemini 4 Argon supports long decode capabilities by sharing a 1 million token limit across input, thinking, and output.
   • Triggers enable scheduled agent executions, such as daily pre-market financial analysis and automated webhooks.
4. 08:50 - 13:12

   Stateful Primitives and Security

   State management is divided into interactions for conversation history and environments for file system persistence. This separation facilitates multi-agent workflows where one agent researches and saves files, and a subsequent agent picks up the environment to perform calculations. Security is maintained by keeping actual credentials out of the sandbox, utilizing an intermediary proxy to handle authentication tokens.

   • Interactions manage conversation turns, while environments maintain the state of the sandbox file system.
   • Separating interactions and environments allows multiple agents to reuse the same file workspace with fresh context windows.
   • An API proxy protects sensitive credentials by injecting dummy tokens into the sandbox and safely adding real secrets during external API calls.
5. 13:12 - 19:36

   Live Demos and Application Integration

   Practical implementations demonstrate how multi-agent systems divide complex pipelines into specific, manageable roles. An SEC researcher agent can extract metrics to a shared file, allowing a quantitative agent to calculate valuations and a reporting agent to compile a PDF summary. These systems integrate smoothly into external apps and can be tested directly within AI Studio.

   • Multi-agent workflows enable specialized agents to fan out tasks, such as SEC filing research followed by quantitative valuation.
   • Managed agents integrate easily into applications to automate data visualization from uploaded documents or receipts.
   • Developers can start building managed agents for free via Google AI Studio and its built-in gallery examples.