Google Updates Gemini API Managed Agents with 3.6 Flash Default, Environment Hooks, and Free Tier
Summary
- • Gemini 3.6 Flash is now the default model for all Gemini API Managed Agents
- • New environment hooks let developers intercept, validate, or block agent tool calls in real time
- • Model selection API added, supporting three Gemini tiers by cost and performance
- • Free tier access now available for Gemini API Managed Agents developers
Details
3.6 Flash now default
antigravity-preview-05-2026 agent now runs Gemini 3.6 Flash by default; existing API users get the upgrade automatically, no code changes needed
Environment hooks system
Hooks config in .agents/hooks.json runs command or HTTP handlers on pre_tool_execution or post_tool_execution events; regex matcher targets specific or all tools
Security gate deny flow
If a pre-execution hook script returns {"decision": "deny", "reason": "..."}, the tool call is skipped and the rejection reason is injected into the model's context window
Three-tier model selection
Developers can pass agent_config.model to select Gemini 3.6 Flash (balanced, default), 3.5 Flash (general agentic), or 3.5 Flash-Lite (lowest cost and latency)
Free tier and MCP continuity
Free tier access added; release extends earlier background tasks and remote MCP server integration, deepening Google's agentic API platform offering
Feature breakdown for Gemini API Managed Agents July 2026 update
What This Means
Google's Gemini API Managed Agents update addresses two enterprise priorities: production-grade safety controls and model cost flexibility. The environment hooks system — allowing custom security gates to intercept and deny agent tool calls in real time — is a meaningful capability for enterprises deploying AI agents in sensitive production workflows. Combined with the free tier addition and configurable model tiers, Google is positioning its managed agents platform to attract both developer experimentation and enterprise-scale deployments, competing directly with similar agentic API offerings from OpenAI and Anthropic.