Tech • IA • Crypto
Getting Started with Managed Agents
8/10TL;DR
Google has introduced managed agents in the Gemini API and AI Studio, enabling developers to build customizable AI agents that can execute code, browse the web, and operate within secure cloud-based environments.
KEY POINTS
Managed Agents Run in Secure Sandboxes
The new system allows AI agents to operate inside a secure Linux sandbox hosted by Google, where they can execute code, manage files, and perform tasks autonomously. This environment isolates operations while still enabling complex workflows such as scripting, data processing, and file generation.
Powered by Gemini 3.5 Flash
The agents are driven by Gemini 3.5 Flash, a model optimized for fast, agentic workflows. It supports reasoning, multi-step execution, and tool usage, making it suitable for coding tasks, automation, and interactive problem-solving.
AI Studio Provides No-Code Entry Point
Developers can quickly experiment via AI Studio, which now includes an “Agents” tab with prebuilt templates. Examples include tools for customer support, data analysis, and repository maintenance, allowing users to launch tasks with minimal setup.
End-to-End Task Automation Demonstrated
In one example, an agent generated a weather dashboard by fetching live data, parsing it with Python, and producing an interactive HTML interface styled with Tailwind CSS. The agent handled the full workflow, from data retrieval to front-end generation, within a single command.
Transparent Execution and File Access
Users can observe each step the agent performs, including command execution and file creation. Outputs such as scripts, HTML files, and visualizations can be downloaded directly from the sandbox, providing visibility and reproducibility.
Customizable Behavior via Sources and Skills
Agents can be tailored using configuration files like agents.md and skills.md, which define behavior, tone, and capabilities. Developers can also attach scripts, datasets, or entire GitHub repositories as sources, enabling highly specialized agents.
API Support for Programmatic Control
The Gemini API includes an interactions endpoint designed for agent workflows. Developers can initialize agents, send tasks, and maintain multi-step conversations using interaction IDs and persistent environments.
Multi-Step and Stateful Workflows
Agents support ongoing sessions where outputs from one step feed into the next. For example, after generating a Fibonacci sequence, an agent can continue by plotting it and saving the result as an image within the same environment.
Streaming and Real-Time Feedback
The API allows streaming responses, enabling developers to display intermediate steps as agents execute tasks. This supports more interactive applications and improved user experience in real-time systems.
File Retrieval via REST API
While SDK support is still evolving, developers can retrieve generated files by calling a REST endpoint to download a snapshot of the sandbox environment, including all created assets such as scripts and visual outputs.
Custom Agent Creation via API
Developers can create fully custom agents using API calls, defining base models, instructions, and capabilities. One example included a technical explainer agent that generates slide decks, complete with structured content and code snippets.
CONCLUSION
Google’s managed agents expand the capabilities of the Gemini ecosystem by combining autonomous execution, customization, and secure infrastructure, positioning them as a powerful tool for building advanced AI-driven workflows and applications.