Managed Agents in the Gemini API

Key Concepts

• Managed Agents: Autonomous agents within the Gemini API that operate in a sandboxed, remote Linux environment.
• Interactions API: A new interface designed to handle complex, multi-step agentic workflows beyond simple content generation.
• Anti-Gravity Agent: The underlying harness and backbone powering the managed agent experience and the Anti-Gravity IDE.
• Agent-Native Development: A paradigm shift where APIs, documentation, and development workflows are designed specifically for AI agents to consume and execute.
• Markdown-Driven Configuration: Using structured Markdown files (agents.md) to define agent behavior, system instructions, and skills.

---

1. Managed Agents in the Gemini API

Managed agents allow developers to trigger autonomous problem-solving capabilities via a single API call.

• Technical Architecture: Agents run in an isolated, remote Linux sandbox. This allows them to execute Bash commands, write code, and create files without exposing the developer's production servers to security risks.
• Powering Models: The initial launch is powered by Gemini 3.5 Flash and the Google "Anti-Gravity" agent.
• Customization: Developers can build custom managed agents by defining system instructions and personal skills, making them easily deployable for internal teams or external customers.

2. The Evolution: From generateContent to Interactions API

The speakers highlighted a shift in how developers interact with frontier intelligence:

• The Old Era: The generateContent API was built for simple, turn-based content generation (User message $\rightarrow$ Model response).
• The New Era: As agentic workflows (function calling, subagents, long-running tasks) became complex, the Interactions API was created. It treats the environment, tools, and multi-step reasoning as first-party principles.
• Data Model: Unlike the previous model, the Interactions API supports a continuous stream of steps (e.g., tool calls, subagent invocations, environment interactions) rather than just a simple user-model exchange.

3. Real-World Application: "Daily Hacker Bites"

To demonstrate the power of managed agents, the team showcased a "radio production" app:

• Process: The agent monitors Hacker News, identifies top topics and opposing viewpoints, writes a script, and generates a three-minute radio show.
• Technical Execution: It utilizes the Interactions API to call the Gemini model for scriptwriting and integrates with Lyria for background music generation.
• Efficiency: A process that involves analyzing 18+ sources and generating complex audio is triggered by a single API call.

4. Developer Experience and "Agent-First" Documentation

The team emphasized making the API accessible to both human developers and other AI agents:

• Agent-Readable Docs: Documentation is now provided in Markdown format, allowing coding agents to search, parse, and understand the API documentation natively.
• MVP Server: A tool provided to help developers connect skills to their coding agents, allowing them to iterate on agent design within the Anti-Gravity environment.
• Markdown Configuration: Agents are defined using prose in agents.md files. The speakers noted that this aligns with the perspective that "English (or Markdown) is the hottest programming language," as it provides the necessary structure for AI to interpret intent.

5. Key Perspectives and Quotes

• On the complexity of agents: "It used to be so simple and now it's very complicated... We want to build an API that is essentially representing these frontier capabilities."
• On the future of coding: The speakers referenced Andrej Karpathy’s sentiment that English is the hardest programming language, suggesting that Markdown is the ideal structured syntax for the future of software development.
• On Flexibility: Developers are not forced into the managed agent ecosystem; they can use the Gemini model via the Interactions API and hook it into any framework of their choosing.

Synthesis and Conclusion

The introduction of managed agents in the Gemini API represents a transition from static content generation to dynamic, autonomous execution. By providing a sandboxed Linux environment and an "agent-first" API, Google is enabling developers to build complex, multi-step applications with minimal infrastructure overhead. The move toward Markdown-based configuration and agent-readable documentation signals a broader industry shift where the primary "users" of developer tools are increasingly becoming the AI agents themselves. Developers can begin experimenting with these tools via the Gemini API documentation and AI Studio.