TL;DR
Google said on July 28, 2026 that managed agents in the Gemini API now run Gemini 3.6 Flash by default, with no code changes required. The update also adds environment hooks that can block, lint, or audit tool calls inside the agent sandbox, plus token budget caps, scheduled triggers, and free tier access for projects without active billing.
Google DeepMind said on July 28, 2026 that managed agents in the Gemini API now run Gemini 3.6 Flash by default, and that the platform is gaining environment hooks, token budget controls, scheduled triggers, and free tier access. The changes, announced by Philipp Schmid and Mariano Cocirio of Google DeepMind, apply to the antigravity-preview-05-2026 agent in the Gemini Interactions API and require no code changes — the next interaction picks up the new model automatically. For developers running autonomous, multi-turn agent workloads, the release pairs a newer default model with new levers to govern and cap what agents do inside their remote sandboxes.
The antigravity-preview-05-2026 agent now defaults to gemini-3.6-flash, which Google describes as its balanced model for reasoning, coding, and tool use. Developers can override the default by passing agent_config.model when creating an interaction: gemini-3.5-flash remains available for general agentic workflows, and gemini-3.5-flash-lite targets the lowest latency and cost in the Gemini 3.5 family. Existing integrations need no changes to pick up the new default, according to the announcement.
The headline feature is environment hooks. By adding a .agents/hooks.json file to an environment, developers can register custom scripts that run on pre_tool_execution or post_tool_execution events inside the agent’s sandbox. Matchers accept regular expressions — targeting code_execution|write_file, for example, or catching every tool with *. A pre-execution hook that returns {“decision”: “deny”, “reason”: “…”} causes the tool call to be skipped, and the rejection reason is fed back into the model’s context, Google said. Hooks also support http handlers that POST to an external endpoint.
On cost and automation, Google said managed agents are now available on free tier projects, letting developers experiment with an API key from a project without active billing. A new max_total_tokens option in agent_config caps total token consumption — input, output, and thinking — for a run. When the cap is reached, execution pauses safely and the interaction returns status “incomplete”; the environment state is preserved, and the task can resume with a fresh budget by passing previous_interaction_id. Scheduled triggers round out the release, automating recurring agent tasks.
Managed Agents Get Gemini 3.6 Flash, Hooks & Budget Guards
Managed agents in the Gemini API now run Gemini 3.6 Flash by default — no code changes required. The update adds environment hooks that can block, lint, or audit tool calls inside the sandbox, plus token budget caps, scheduled triggers, and free tier access.
The next interaction picks up Gemini 3.6 Flash automatically on existing integrations.
pre_tool_execution and post_tool_execution scripts run inside the agent’s own sandbox.
Managed agents now work with an API key from a project without active billing.
Five New Levers For Autonomous Agent Workloads
Managed agents execute multi-turn loops — reasoning, code execution, package installs, file management, and web retrieval — inside an isolated cloud sandbox from a single API call. This release pairs a newer default model with governance controls that live inside the execution path itself.
Gemini 3.6 Flash By Default
Google’s balanced model for reasoning, coding, and tool use. Override any time via agent_config.model when creating an interaction.
Environment Hooks
Register scripts via .agents/hooks.json that run on tool-call events inside the sandbox — block, lint, or audit before anything executes.
Token Budget Caps
max_total_tokens caps input, output, and thinking tokens per run. At the cap, execution pauses safely and returns status “incomplete”.
Scheduled Triggers
Automate recurring agent tasks on a schedule — rounding out a release aimed at hands-off, production-grade agent operations.
Free Tier Availability
Experiment with managed agents using an API key from a project without active billing — a practical on-ramp before committing spend.
Resume With Fresh Budget
Overruns preserve environment state. Pass previous_interaction_id and the task resumes with a fresh token budget — no lost work.
Three Models, One Config Field
The default applies automatically. For explicit selection, pass the model identifier in agent_config.model when creating the interaction.
| Model Identifier | Positioning | Reasoning & Tools | Latency / Cost | Best For |
|---|---|---|---|---|
| gemini-3.6-flashNew Default | Balanced flagship of the release | ✓ Strong | ~ Balanced | Reasoning, coding, and tool use in production agent loops |
| gemini-3.5-flash | Previous-generation workhorse | ✓ Solid | ~ Balanced | General agentic workflows already tuned on the 3.5 family |
| gemini-3.5-flash-lite | Speed-first lightweight tier | ~ Adequate | ✓ Lowest | High-volume tasks targeting the lowest latency and cost |
Environment Hooks: Governance Inside The Sandbox
Until now, the sandbox being remote meant customer-side validation code had nowhere to run. Hooks execute developer-supplied logic inside the sandbox itself — moving governance from wrapper code around the API into the agent’s execution path.
Agent Plans Tool Call
Model decides to invoke a tool: code_execution, write_file, web retrieval…
Matcher Fires
Regex matchers target tools — e.g. code_execution|write_file, or * for everything.
Pre-Execution Hook
Script runs on pre_tool_execution. An http handler can POST to an external endpoint.
Decision
Returning {“decision”:”deny”,”reason”:”…”} skips the call entirely.
Feedback Loop
The rejection reason is fed back into the model’s context; post_tool_execution audits results.
// pre_tool_execution hook response { "decision": "deny", "reason": "write_file outside approved paths is blocked by policy" } // → tool call skipped · reason injected into model context
Where Each Lever Lands
An editorial read of how the update’s components trade off capability against overhead — from zero-touch defaults to hands-on governance.
Why Hooks Matter
Google said teams are already building production-grade validation pipelines with hooks — naming OffDeal, whose agent “Archie” assembles company decks, as the example.
Before agent hooks, we couldn’t do this on Gemini’s managed agents: the sandbox is remote, so our validation code had nowhere to run.
With hooks, a post_tool_execution hook triggers our pipeline inside the sandbox — fetching candidates, enforcing pixel-level quality checks, verifying each logo with Gemini vision, and publishing a manifest of approved files.
From Announcement To Developer Action
Jul 28, 2026
DeepMind announcement goes live
3.6 Flash
Next interaction auto-upgrades
hooks.json
Add .agents/ scripts to sandbox
Budget Cap
Overrun → “incomplete”, resumable
Triggers
Recurring tasks on schedule
Free Tier
Try it without active billing
What Google Hasn’t Said
GA timeline — no date for antigravity-preview-05-2026 leaving preview.
Pricing & rate limits — nothing beyond free tier availability.
Trigger syntax — full scheduled-trigger configuration not spelled out.
Hook failure semantics — e.g. script timeouts deferred to separate docs.
Developer FAQ
Do I need to change my code to get 3.6 Flash?
No. The agent picks it up automatically on your next interaction. To choose a different model, pass agent_config.model.
What can environment hooks actually block?
Any tool call in the sandbox. A pre-execution hook matches tools by regex and returns a deny decision; the call is skipped and the reason enters the model’s context.
Are managed agents free to try now?
Yes. They’re available on free tier projects — experiment with an API key from a project without active billing.
New Guardrails for Autonomous Agent Workloads
Managed agents execute multi-turn autonomous loops — coordinating reasoning, code execution, package installation, file management, and web retrieval inside an isolated cloud sandbox from a single API call. Until now, the sandbox being remote meant customer-side validation code had nowhere to run. Hooks change that by executing developer-supplied logic inside the sandbox itself, moving governance from wrapper code around the API into the agent’s execution path. Google said teams are already using hooks to build production-grade validation pipelines, though it named only one customer, OffDeal, in the announcement.
The budget cap addresses a known failure mode of agentic systems: runaway loops that consume large token budgets before anyone notices. Returning a structured “incomplete” status with preserved state makes overruns recoverable rather than lost work. Free tier access, meanwhile, lowers the barrier for individual developers to test the Interactions API before committing billing — a practical on-ramp for a capability Google has so far kept behind paid projects.
AI development environment hooks tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Background Tasks to Deeper Sandbox Control
The July 28 release builds on Google’s previous managed agents update, which introduced background tasks and remote MCP server integration. Managed agents are accessed through the Gemini Interactions API, where a single API call coordinates reasoning, code execution, package installs, file management, and web retrieval in an isolated cloud environment. Google offers SDKs such as @google/genai for TypeScript and JavaScript, with Python and cURL examples in its agent documentation. The agent remains in preview, as reflected in its antigravity-preview-05-2026 identifier. Google also published a companion skill — installable via npx skills add google-gemini/gemini-skills — that gives AI coding assistants access to the Interactions API.
“Managed Agents in Gemini API are getting environment hooks, model selection, and free tier access.”
— Google DeepMind announcement, authored by Philipp Schmid and Mariano Cocirio
Patriola's Guide to Claude: Token Budgets: Control What Your Claude Sessions Actually Cost
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What Google Has Not Yet Detailed
Several specifics were absent from the July 28 announcement. Google did not say when the antigravity-preview-05-2026 agent will reach general availability, nor did it detail pricing or rate limits for managed agent usage beyond free tier access. The full configuration syntax for scheduled triggers was not spelled out in the release materials, and failure-handling semantics for hooks — such as what happens when a hook script times out — are deferred to separate documentation. The OffDeal deployment is a customer testimonial; Google said “teams” are using hooks in production but named no others.
Building Agentic AI with n8n: Design No-Code AI Agents and Workflow Automations Using LLMs, APIs, and Event-Driven Triggers
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Developers Can Start Using the Update
The Gemini 3.6 Flash default applies automatically — the next interaction uses it with no code changes. Developers wanting hooks can add a .agents/hooks.json file and hook scripts to their environments; Google’s hooks documentation covers HTTP handlers and failure handling. The @google/genai npm package exposes the new agent_config fields, and Google’s documentation lists model identifiers for explicit selection. Milestones to watch include the agent’s move out of preview and any additions to hook or trigger capabilities in subsequent releases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Do I need to change my code to get Gemini 3.6 Flash?
No. Google said the antigravity-preview-05-2026 agent picks up Gemini 3.6 Flash automatically on your next interaction. To use a different model — such as gemini-3.5-flash-lite for lower cost — pass agent_config.model when creating the interaction.
What can environment hooks actually block?
Any tool call the agent makes inside its sandbox. A pre_tool_execution hook can match tools by regular expression and return a deny decision with a reason; the call is then skipped and the reason is added to the model’s context, according to Google.
Are managed agents free to try now?
Yes. Google said managed agents are available on free tier projects, so developers can experiment using an API key from a project without active billing. Usage limits for the free tier were not detailed in the announcement.
What happens when my agent hits its token budget?
Execution pauses safely and the interaction returns status “incomplete”. The environment state is preserved, and you can continue the task with a fresh budget by passing previous_interaction_id in a new call.
Which models can managed agents use as of July 28, 2026?
Gemini 3.6 Flash (the new default, positioned for balanced reasoning, coding, and tool use), Gemini 3.5 Flash (the previous generation), and Gemini 3.5 Flash-Lite (the lowest latency and cost in the 3.5 family), according to Google.
Source: Google AI
