Google has rolled out new capabilities for Managed Agents in the Gemini API. These updates help developers build reliable, production-ready agents. The highlights include a new default model, environment hooks, budget controls, scheduled triggers, and free tier access.
Takeaways 🚀
✅Default model now uses Flash
✅Hooks control sandbox tool behavior
✅Budgets help prevent runaway costs
✅Scheduled triggers automate recurring tasks
✅Free tier broadens developer access
Let’s break down what’s new.
What Are Managed Agents?
Managed Agents live inside the Gemini Interactions API. With a single API call, an agent can handle several tasks at once. It reasons, runs code, installs packages, manages files, and retrieves web content. All of this happens inside an isolated cloud sandbox.
This setup keeps your workflow simple. You don’t need to stitch together multiple tools or services.
Gemini 3.6 Flash Is Now the Default
Managed Agents now run on Gemini 3.6 Flash by default. You don’t need to change your code. Your next interaction picks up the new model automatically.
Gemini 3.6 Flash offers a strong balance across reasoning, coding, and tool use. But you can pick a different model if you prefer. Just pass agent_config.model when you create an interaction.
Here are the supported models:
- Gemini 3.6 Flash (
gemini-3.6-flash, default): A balanced model for reasoning, coding, and tool use. - Gemini 3.5 Flash (
gemini-3.5-flash): The previous generation for general agentic workflows. - Gemini 3.5 Flash-Lite (
gemini-3.5-flash-lite): The lowest latency and cost in the Gemini 3.5 family.
Want lower costs? Choose Gemini 3.5 Flash-Lite. Want to lock in a specific version? Simply pin your model of choice.
Environment Hooks: Block, Lint, and Audit Tool Calls
Environment hooks give you more control over what an agent does inside its sandbox. You can run your own scripts before or after every tool call.
To set this up, add an .agents/hooks.json file to your environment. The runtime then runs your handlers on two events: pre_tool_execution and post_tool_execution.
How Hooks Work
You can target specific tools using the matcher field. It supports regular expressions. Use | to match multiple tools, or * to catch everything.
Here’s what hooks let you do:
- Block risky actions. Run a script before a tool call. If the script returns a “deny” decision, the agent skips that call. It also passes the reason back to the model.
- Enforce code style. Run a linter after every tool finishes. This keeps your code clean and consistent.
- Audit activity. Track and review what the agent does inside the sandbox.
Hooks also support HTTP handlers. These post directly to an external endpoint, so you can connect your own services.
A Real-World Example
Teams already use hooks to build strong validation pipelines. OffDeal, an AI-native investment bank, is one example.
Their AI analyst, Archie, builds banker-ready decks. These decks often need 30 or more company logos. Each logo must meet strict quality standards. It needs the right size, aspect ratio, transparent background, and high contrast.
Before hooks, OffDeal couldn’t run these checks. The sandbox is remote, so their validation code had nowhere to run. Now, a post_tool_execution hook triggers their pipeline the moment Archie writes a company list. The pipeline fetches logos, runs pixel-level quality checks, and verifies each one with Gemini vision. Only approved images make it into the final deck.
Cost Control and Automation Features
New features help you manage spending and automate repeat tasks.
Free Tier Access
Managed Agents now work on free tier projects. You can test agentic workflows using an API key from a project without active billing. This makes it easier to experiment before you commit.
Budget Controls
Managed Agents run multi-turn loops on their own. Complex tasks can burn through a lot of tokens. To keep costs in check, you can set a limit with max_total_tokens inside agent_config. This caps the total token use, including input, output, and thinking.
When the agent hits the limit, it pauses safely. The interaction returns a status: "incomplete". Your environment state stays intact. You can then continue where the agent stopped by passing previous_interaction_id with a fresh budget.
Scheduled Triggers
Triggers let you automate recurring tasks. A trigger ties together an agent, an environment, a prompt, and a cron schedule. It then fires on its own, with no manual work needed. Each run reuses the same sandbox, so your files stay in place across runs.
Environments API
The Environments API gives you direct control over your sandbox sessions. You can list, inspect, and delete them from code. This helps you recover environment IDs after a disconnect. You can also clean up sandboxes when your pipeline finishes, instead of waiting for the 7-day time limit.
Get Started With Managed Agents
These updates turn Managed Agents into flexible, cost-controlled workers. They run on their own inside real development environments. And they do it without breaking your budget or needing outside orchestration.
To learn more, explore the Gemini Interactions API overview and the Managed Agents quickstart. There you can dig into custom agent definitions, environment settings, network rules, and advanced streaming patterns.
