Tech • IA • Crypto
Managed Agents in the Gemini API
8/10TL;DR
Google has introduced managed agents in the Gemini API, enabling developers to deploy autonomous AI agents with a single call in a secure, sandboxed environment.
KEY POINTS
Single-call autonomous agents
The new managed agents capability allows developers to invoke an autonomous agent with a single API call. These agents can independently execute complex tasks such as writing code, running commands, and orchestrating workflows. The goal is to simplify access to advanced AI behaviors that previously required significant engineering overhead.
Sandboxed Linux execution environment
Each agent operates inside a remote, isolated Linux sandbox, where it can safely execute code, create files, and run shell commands. This design removes the need to expose production systems to AI-generated code, reducing security risks while enabling powerful automation.
Powered by Gemini 3.5 Flash and Anti-Gravity
The system is initially powered by Gemini 3.5 Flash alongside a new Anti-Gravity agent framework, which serves as the backbone for agent orchestration. The same framework also underpins development tools like the Anti-Gravity IDE, ensuring consistency across environments.
Customizable and reusable agents
Developers can build custom managed agents by defining system instructions, adding specialized “skills,” and packaging them for reuse. These agents can then be deployed internally or exposed to customers, supporting use cases ranging from internal automation to full-scale applications.
Integration via the Interactions API
Managed agents are accessed through the Interactions API, a newer interface designed to unify communication with both models and agents. This replaces earlier paradigms focused purely on content generation, reflecting a shift toward multi-step, tool-driven AI workflows.
From simple prompts to multi-step workflows
Unlike traditional chat-based APIs, the new system supports continuous streams of actions rather than turn-based exchanges. Agents can perform sequences of steps such as tool calls, sub-agent delegation, and reasoning chains before returning results, enabling more complex and autonomous behavior.
Built-in tools and first-class environment support
Tools, execution environments, and intermediate steps are treated as first-class elements within the API. This allows agents to seamlessly integrate capabilities like function calling, file generation, and external tool usage without requiring custom infrastructure.
Demo: automated radio show generation
A demonstration application showcased an agent that transforms trending discussions into a three-minute radio-style program. The agent aggregates sources, generates a script, produces audio segments, and even creates music using Lyria, highlighting the system’s ability to coordinate multiple AI capabilities in one workflow.
Agent-first developer experience
The platform introduces an “agent-native” development approach, where agents can be defined using simple Markdown files. Developers specify behavior in an agents.md file and define skills similarly, making agent creation more accessible and aligned with familiar documentation practices.
Documentation optimized for AI usage
Documentation has been redesigned to be machine-readable and easily navigable by agents. Features such as searchable Markdown docs and dedicated integration tools allow coding agents to understand and implement the API more effectively, even with knowledge cutoffs.
Flexible adoption for developers
Developers can choose between fully managed agents or continue using raw models with their own frameworks. This flexibility allows gradual adoption, catering to both simple applications and more advanced, agent-driven systems.
CONCLUSION
Managed agents in the Gemini API mark a shift toward fully autonomous, tool-using AI systems, combining ease of use with powerful execution capabilities in a secure environment.