Managed agents move toward controlled automation
Google has expanded Managed Agents in the Gemini API with a new default model, environment hooks, cost controls, scheduled execution and free tier access. The update is aimed at developers building agentic workflows that need to run code, manage files, install packages and retrieve web content inside an isolated cloud sandbox.The central change is not a single feature, but a broader move to make managed agents easier to govern. Google describes the system as part of the Gemini Interactions API, where one API call can coordinate reasoning, code execution, package installation, file management and web retrieval in a remote environment. The latest additions focus on model choice, validation, budgeting and recurring automation.
Gemini 3.6 Flash becomes the default model
Google says the antigravity-preview-05-2026 agent now runs Gemini 3.6 Flash by default. Developers do not need to change code for that default to apply, according to the post; the next interaction picks it up automatically.The update also adds explicit model selection through agent_config.model when creating an interaction or managed agent. Google lists Gemini 3.6 Flash as the default balanced model for reasoning, coding and tool use. It also lists Gemini 3.5 Flash for general agentic workflows and Gemini 3.5 Flash-Lite, which Google describes as the lowest latency and cost option in the Gemini 3.5 family.
That model selection matters because managed agents can run multi-step autonomous tasks. A development team may want a stronger default for coding and tool use, while another workflow may prioritize lower latency or lower cost. The post frames model pinning as a way to choose the right tradeoff instead of relying only on the default.
Hooks add pre-call and post-call controls
The most governance-focused addition is environment hooks. Google says hooks let developers run custom scripts before or after every tool call an agent makes inside its sandbox. Developers can add a .agents/hooks.json file to an environment, and the runtime executes handlers on pre_tool_execution or post_tool_execution events.The matcher field supports regular expressions, including targeting multiple tools with a pipe character or catching everything with an asterisk. In Google’s example, a security-gate group runs a Python script before code_execution or write_file calls. If that script returns a deny decision with a reason, the tool call is skipped and the rejection reason is passed back into the model’s context.
Post-call hooks can be used for cleanup or validation. Google gives an auto-format example in which a script runs after every tool completes in order to enforce code styling. The post also says hooks can use HTTP handlers that post directly to an external endpoint, with complete HTTP hook definitions and failure-handling details available in the related documentation referenced by Google.
The practical effect is that teams can insert policy checks inside the remote sandbox rather than waiting until after the agent has completed its work. That is important for workflows where the agent may write files, run code or generate artifacts that need immediate review.
OffDeal example shows sandbox validation use case
Google cites OffDeal, described in the source as an AI-native investment bank, as an example of how hooks are being used. Alston Lin, Founder and CTO of OffDeal, said the company’s Archie AI analyst is used by bankers and that banker-ready decks require company logos across buyer tables, sponsor columns and tombstone grids.According to Lin’s statement in the post, those logos often number more than 30 in a single deck and must match the right company, size, aspect ratio, name, transparent background and contrast requirements. He said that before agent hooks, OffDeal could not run its validation code on Gemini managed agents because the sandbox was remote.
With post_tool_execution hooks, OffDeal’s pipeline can run inside the sandbox after Archie writes its company list, according to the post. The process fetches candidates, applies image quality checks, verifies each logo with Gemini vision and publishes a manifest of approved files that are the only images allowed into the deck. The example is narrow, but it illustrates why hook placement matters: validation happens where the agent’s files and outputs are being produced.
Budgets, free tier access and scheduled triggers
Google also added features aimed at controlling cost and automating recurring work. Managed agents are now available on free tier projects, allowing developers to experiment with agentic workflows using an API key from a project without active billing.Because managed agents can execute multi-turn autonomous loops, Google warns that complex tasks can consume significant token budgets. Developers can now pass max_total_tokens inside agent_config to cap total consumption across input, output and thinking tokens. When the agent reaches that limit, execution pauses safely and the interaction returns status incomplete. Google says the environment state is preserved, so a developer can continue where the task stopped by passing previous_interaction_id with a fresh budget.
Scheduled triggers are another automation feature. A trigger binds an agent, environment, prompt and cron schedule into a persistent resource that fires without manual intervention. Each run reuses the same sandbox, which means files persist across executions. Google also says the Environments API lets developers list, inspect and delete sandbox sessions from code, recover environment IDs after a disconnect, or clean up sandboxes when a pipeline finishes rather than waiting for the seven-day TTL.
Why the update matters for developers
The update points to a more operational version of managed agents. Earlier agent demos often focused on autonomous task completion, but production use requires controls around what an agent can do, how much it can spend, when it should run and how teams can inspect or stop its environment.Environment hooks address the policy layer by allowing teams to block, lint or audit tool calls. Budget controls address runaway consumption. Scheduled triggers turn agents into recurring workers, and free tier access lowers the entry point for experimentation. Google’s post presents these changes as part of a path toward cost-controlled, scheduled workers that can operate inside real development environments without external orchestration.
For developers already testing the Gemini Interactions API, the key question is likely not whether agents can run more tasks, but whether those tasks can be bounded and checked. This release gives teams more places to put those boundaries.
Sources
Editorial Team - CoinBotLab