Build Your Own Team of Managed Agents - Jack Wotherspoon, Google DeepMind
Summary
Gemini Managed Agents allow developers to run sophisticated coding agents in secure, scalable cloud sandboxes via a single API call, eliminating the need for complex local infrastructure setup. This architecture enables robust, multi-agent workflows—such as sequential financial research and reporting—by sharing a persistent environment state across different agents. The system is built around the Interactions API and includes advanced features like secure credential proxies and scheduled triggers.
Key takeaways
-
Cloud Execution for Scalability and Security
Running agents in a cloud sandbox (Google Cloud infrastructure) solves local machine limitations, such as scaling constraints and browser interference. A dedicated proxy layer ensures that sensitive credentials (e.g., GitHub tokens) never enter the sandbox, only being injected when needed for specific API calls.
-
Simplified Development via Interactions API
5:26
The Interactions API is the new agent-first SDK, replacing older model-input/output structures. It allows agents to perform multi-step actions, including writing files and calling custom MCP tools, all through a single API call.
-
Multi-Agent Workflows through Shared Environments
17:26
Managed agents support complex workflows by allowing multiple agents to share the same environment ID (the sandbox file system). This enables one agent to perform research and save files, which a subsequent, different agent can then pick up and build upon, maintaining context and state.
Technical details
-
Managed Agent Architecture
522s
Managed agents utilize a cloud sandbox and the Anti-Gravity harness. Key building blocks include: the **Harness** (Anti-Gravity, managing tool calls); the **Environment** (the stateful file system in the cloud); **Skills** (uploadable custom logic); **Credentials** (secure injection of OAuth/tokens via a proxy); and **Triggers** (allowing scheduled execution or webhooks).
-
Advanced Model Capabilities
702s
The platform supports tuning the agent using different models (e.g., Gemini Flash for speed, Gemini 4 Argon for deep reasoning/finance). Gemini 4 also features a long decode capability, expanding the context window up to 1 million tokens, which is shared across input, thinking, and output.
-
Interactions and Environments Primitives
1046s
These two core primitives manage state: **Interactions** manage the conversation history (input/output turns), while **Environments** manage the actual file system state within the sandbox. Passing the Environment ID between agents allows for stateful, multi-step processes.
Mentioned resources
- AI Studio
- Gemini API
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.