πŸ“° Key Highlights

AI news flash: Gemini API’s Managed Agents feature rolls out three updates: model upgrade, environment hooks, and free tier access. This builds on the previously launched background tasks and remote MCP server integration. Through the Gemini Interactions API, a single API call can orchestrate reasoning, code execution, package installation, file management, and web fetching, all inside an isolated cloud sandbox. Developers can run npx skills add google-gemini/gemini-skills --skill gemini-interactions-api to let their AI coding assistant pick up the Interactions API skill.

On the model side, the antigravity-preview-05-2026 agent now defaults to Gemini 3.6 Flash β€” no code changes needed, it’ll kick in automatically on your next interaction. You can also pin a model manually via the agent_config.model parameter, for example to switch to the lower-cost Gemini 3.5 Flash-Lite. Three models are currently supported: Gemini 3.6 Flash (default β€” balanced reasoning, code, and tool use), Gemini 3.5 Flash (previous-generation general-purpose agent model), and Gemini 3.5 Flash-Lite (the lowest latency and lowest cost option in the 3.5 series).

Environment hooks let you run custom scripts before and after every tool call inside the sandbox. Just drop a .agents/hooks.json config file into your environment, and the execution environment will fire the matching handler on pre_tool_execution or post_tool_execution events. The matcher field supports regex, so you can match multiple tools with | or all tools with *. The example shows two hook groups: security-gate runs a gate.py script for security review before every code_execution or write_file call β€” if the script returns a deny decision, that tool call is skipped entirely; auto-format runs an auto-formatter script after every tool call. Full details in the original link.

πŸ’¬ JudyAI Lab Perspective

This round of Gemini API Managed Agents ships three updates at once β€” model upgrade, environment hooks, and free tier access β€” pushing what was once a plain inference service toward a more configurable, governable agent infrastructure.

What we think is really worth paying attention to isn’t the default model swap to Gemini 3.6 Flash, but the design of environment hooks: drop a config file in, and you can inject custom scripts before and after every tool call, use regex to decide which actions to intercept, and even let a security script veto a code execution or file write outright. This means the center of gravity for agent systems is shifting from “how to write prompts” to “how to design checkpoints in the execution flow” β€” and that’s more worth thinking through clearly than picking a model.

If you’re wiring up a similar agent API, it’s probably worth listing first: which tool calls do you absolutely need to be able to intercept or veto?

πŸ“… Source Info

πŸ”— Further Reading