Gemini API Managed Agents: Antigravity Agent 09-2026 Replaces the May Version on October 5

On September 17, 2026, Google released a new version of the Antigravity agent, the general-purpose agent behind Managed Agents in the Gemini API. The Gemini API release notes state that the new version, antigravity-preview-09-2026, replaces and deprecates antigravity-preview-05-2026, and give October 5, 2026 as the shutdown date for the older version.

Managed Agents let a developer start an AI agent with a single API call. Google hosts the environment where the agent runs code, edits files and browses the web. For companies adding agent features to a product, often on a Node.js development stack or a similar backend, this removes much of the infrastructure work. It also means that Google's release schedule becomes part of your own.

This update is a clear example. The service is still in public preview, the new version changes how the agent's tools are called, and the gap between release and shutdown of the old version was under three weeks.

What changed on September 17, 2026

According to the release notes entry for that date, the new agent version changes the built-in tools the agent uses inside its sandbox. Tool parameters now use PascalCase naming where they previously used snake_case. File edits are made as line-range replacements instead of full file rewrites, and file reading uses line-based parameters instead of byte offsets. The entry also lists two added search tools, find_by_name and grep_search, and says shell execution and web search are unchanged.

Google's Antigravity agent documentation describes the current version as built with Gemini 3.8 Flash, the model Google made generally available on September 2, 2026. Developers can select a different model, and the documentation lists Gemini 3.5 Flash-Lite as the option for low-latency and cost-sensitive tasks.

For a business owner these sound like small technical details, and for many integrations they will be. They matter when your own code inspects, logs, filters or blocks the agent's tool calls, because that code refers to the tool parameters by name.

The October 5, 2026 shutdown

Google's deprecations page lists antigravity-preview-05-2026 with a release date of May 19, 2026, a shutdown date of October 5, 2026 and antigravity-preview-09-2026 as the replacement. Any application that still names the May version in its requests should be expected to fail from that date onward, so a reference left in place after October 5, 2026 is a fault to fix, not a migration to schedule.

Because the agent version is written into the request as an identifier, it usually sits in a configuration file or directly in code. Finding and updating it is quick. Confirming that everything around it still behaves correctly takes longer.

A quick recap of Managed Agents

Google introduced Managed Agents in a blog.google post on May 19, 2026, as a public preview available through the Interactions API and Google AI Studio, with an enterprise preview on the Gemini Enterprise Agent Platform. The post explains that developers can customize the agent with instructions and skills written as Markdown files.

The Managed Agents overview describes the environment: an isolated Linux sandbox, Ubuntu-based with Python 3.12 and Node.js 22, that shuts down when idle and restores its state on the next request. The same page says an environment is permanently deleted after seven days of inactivity and that an account can hold up to 1,000 managed agents.

Google has added features steadily since launch. A July 7, 2026 post announced background execution for long-running tasks, connections to remote Model Context Protocol (MCP) servers, custom functions and credential refresh. A July 28, 2026 post added environment hooks, token budget controls, scheduled triggers and free tier access.

Cost and security details to understand

Billing

The overview page says Managed Agents are billed on a pay-as-you-go basis for the model tokens and tools the agent uses, and that the compute for the environment is not billed during the preview. It also states that a single interaction typically uses between 100,000 and 3 million tokens. That is a wide range, and it is far more than a normal chat request. The July 28 post describes a setting, max_total_tokens, that pauses an agent when it reaches a token limit while keeping its environment so the task can be continued.

Network access and credentials

The overview page states that the sandbox has unrestricted outbound network access by default and that this can be narrowed with an allowlist of permitted domains. It advises storing secrets as managed credentials, using least-privilege service accounts and short-lived tokens, using only tools from trusted sources, and verifying outputs before relying on them, especially for tasks that modify data.

Hooks

The July 28 post explains that environment hooks run a developer's own scripts before or after each tool call inside the sandbox, and that a hook can deny a call. This is the mechanism for rules such as blocking certain commands or recording every action for audit. It is also the code most exposed to the renamed tool parameters in the September version.

What this means for your business

The points below are our interpretation and are offered as guidance.

A preview service moves quickly and can change in ways that break your code at short notice. That is a reasonable trade for internal tools, prototypes and back-office automation, where a failed run is an inconvenience. It is a weaker fit for a customer-facing feature with uptime commitments, unless someone is responsible for watching Google's release notes and acting on them.

Managed Agents also shift the build-or-buy decision. Hosting your own agent sandbox means building isolation, state handling and tool plumbing yourself. Using Google's removes that work and ties you to Google's models, limits and timetable. Neither is wrong, but the dependency should be a deliberate choice.

It is also worth asking whether a task needs an autonomous agent at all. Many business processes are served better and more cheaply by a fixed workflow with one or two model calls. Our article on AI agents, chatbots and workflow automation sets out how to tell the difference.

What to do next

• Search your code and configuration for antigravity-preview-05-2026 and replace every reference with the September version. From October 5, 2026, the shutdown date Google lists, requests that still name the May version should be expected to fail.
• Review any hooks, logging or parsing code that reads tool calls, and update parameter names and file-edit handling to match the new version.
• Re-run scheduled triggers and long-running tasks in a test environment and compare results and token use.
• Set a token budget on every agent, restrict outbound network access to the domains the agent needs and move any inline secrets into managed credentials.
• Assign one person to follow the Gemini API release notes and deprecations page.

Conclusion

Google released antigravity-preview-09-2026 on September 17, 2026, built with Gemini 3.8 Flash and with changed tool parameters, and set October 5, 2026 as the shutdown date for the previous agent version. Managed Agents in the Gemini API remain in public preview, so further changes of this kind should be expected.

If you already use the service, check that no reference to the May version remains and retest carefully on the September version. If you are evaluating it, weigh the saved infrastructure work against the pace of change. Entrant Technologies builds websites, web applications, mobile apps and custom software, and you are welcome to contact us if you want a second opinion on an agent project.