Alphabet expands Gemini API Managed Agents with Gemini 3.6 Flash defaults, sandbox hooks, and cost controls

Google is rolling out a new set of updates to its Gemini API Managed Agents, aiming to make autonomous “agentic” workflows more dependable in real software development. In a July 28 post, Google said Managed Agents will now default to Gemini 3.6 Flash, and it introduced environment hooks that allow developers to run custom checks around every tool call the agent makes inside a cloud sandbox.

Managed Agents in the Gemini Interactions API are designed so a single API call can coordinate an end-to-end sequence, including reasoning, code execution, package installation, file management, and web retrieval, all within an isolated environment. That sandboxing approach is intended to contain the agent’s actions and reduce the need for external orchestration scripts that developers typically build themselves.

A centerpiece of the update is environment hooks, which let developers insert custom code that runs before or after the agent’s tool calls. Developers create a file named agents/ in their environment configuration, then define handlers tied to events such as pre_tool_execution and post_tool_execution. Google said hooks use a “matcher” field that supports regular expressions, letting developers target specific tools or capture all tool calls.

Google also described new operational controls intended to address two common challenges with autonomous agents: cost predictability and reliability over longer runs. To prevent runaway multi-turn loops that can spend large amounts of compute, developers can set max_total_tokens in agent_config to cap total consumption, covering input, output, and “thinking.” If the agent hits the limit, execution pauses and returns an “incomplete” status, with environment state preserved so developers can continue by passing a previous_interaction_id with a fresh budget.

Beyond spending caps, Google said it added scheduled triggers that bind an agent, its environment, prompt, and a cron schedule into a persistent resource. Each run can reuse the same sandbox, with files persisting across executions. In a related capability, Google noted that the Environments API lets developers list, inspect, and delete sandbox sessions from code, including the ability to recover environment IDs after a disconnect and clean up sandboxes instead of waiting for a stated time-to-live.

Google’s update also emphasizes that the new defaults and tools are meant to require minimal migration. For the antigravity-preview-05-2026 agent, Google said it now runs Gemini 3.6 Flash by default with “no code changes required,” and that the next interaction will automatically pick up the change. Developers can also explicitly select models by passing agent_config.model when creating an interaction or managed agent, including an option called Gemini 3.5 Flash-Lite for lower cost or “pinning” to a preferred model.

In the post, Google provided an example of how hooks can solve validation problems that appear when agent work happens inside a remote sandbox. It cited Offdeal, an “AI-native investment bank,” describing a workflow for its AI analyst, Archie, where a requirement for banker-ready decks is having correct company logos with strict formatting rules. Google said that prior to agent hooks, its validation code had nowhere to run because the sandbox was remote, but with post_tool_execution hooks, Offdeal’s pipeline can verify logos inside the sandbox at the moment Archie produces its company list, enforce pixel-level quality checks, verify images with Gemini vision, and produce a manifest of approved files for inclusion in decks.

Google also highlighted that Managed Agents are now available on free tier projects, allowing developers to experiment using an API key from a project without active billing. It framed the combination of defaults, hooks, budget controls, and recurring scheduling as a way to turn managed agents into “cost-controlled, scheduled workers” that operate autonomously inside development environments without breaking budgets or requiring external orchestration.

Still, Google did not provide detailed performance metrics, pricing for the managed agent features beyond the availability on free tier, or any quantitative benchmarks showing how often hooks improve outcomes. It also did not disclose limits around how complex workflows can be before hitting token caps or how consistently scheduled triggers handle failures across long periods, leaving those operational characteristics to developer testing and documentation.

What to watch next is how quickly developers adopt environment hooks for compliance-style validation and how often model selection and budget caps become part of standard agent templates. Google’s reference to recurring workers and sandbox lifecycle management suggests it wants Managed Agents to move from demos toward production pipelines, where controls and repeatability matter as much as raw model capability.