Google upgrades Gemini API agents with Gemini 3.6 Flash and automation
In brief
- Gemini 3.6 Flash now default model for API managed agents, reducing reasoning steps and tool calls
- Environment hooks enable developers to inspect, block, or audit agent actions before execution
- Scheduled task triggers allow recurring agent workflows without separate orchestration infrastructure
- Token budget controls pause execution when limits reached, preventing runaway consumption
- Managed agents available on free-tier Gemini API projects for developer testing
Faster, more efficient agentic workflows
Gemini 3.6 Flash is designed for coding, reasoning, multimodal analysis, and repeated tool use. Google says the model completes agentic workflows with fewer reasoning steps, tool calls, and debugging loops than Gemini 3.5 Flash. Developers can still select Gemini 3.5 Flash and Gemini 3.5 Flash Lite as alternative models through agent configuration if needed.
Managed agents coordinate reasoning, code execution, package installation, file operations, and web retrieval through a single Gemini Interactions API call. Each agent operates inside an isolated cloud sandbox managed by Google.
New execution controls and safety hooks
The update introduces environment hooks allowing developers to execute custom scripts before or after an agent uses a tool. Pre-execution hooks can inspect and block actions such as running code or writing files. Post-execution hooks can automatically lint generated code, audit outputs, verify files, or send results to an external endpoint.
AI investment banking platform OffDeal is already using this feature. OffDeal uses environment hooks to verify company logos before its agent places them into presentation materials.
Google also added token budget controls for autonomous tasks allowing developers to set a maximum total token limit. When an agent reaches the token limit, execution pauses with an incomplete status rather than continuing to consume tokens.
Scheduled execution and sandbox persistence
Managed agents can now run automatically through scheduled triggers. Developers can connect an agent, prompt, environment, and cron schedule to create recurring workflows without operating a separate orchestration service. Each scheduled run can reuse the same sandbox, meaning files and other environment data can remain available between executions.
A new Environments API lets developers list, inspect, and delete sandbox sessions. The Antigravity preview agent now uses Gemini 3.6 Flash automatically, with no code changes required from developers.
Accessibility expanded too. Managed agents are now available through free tier Gemini API projects, allowing developers to test agentic workflows at no cost. The update builds on the May launch of managed agents, which introduced Google's Antigravity agent, background execution, custom agent configurations, and remote Model Context Protocol server support.
