Antigravity Agent 09-2026: New File Tools, PascalCase Parameters, and an October 5 Shutoff

Google released Antigravity Agent 09-2026 on September 17, 2026, and the agent it replaces, antigravity-preview-05-2026, shuts down on October 5. If your code calls the managed Antigravity agent in the Gemini API, you have days, not weeks, to switch the agent string and test.

For many teams the switch is one line. For others it is a real code change, because the built-in file tools were renamed and their parameters now use PascalCase. This piece walks through what changed, who needs more than a string swap, what stayed the same, what we could not confirm, and a checklist to finish before the shutoff.

The short version: if you run on Google’s remote sandbox and only read the agent’s final text, change antigravity-preview-05-2026 to antigravity-preview-09-2026 and re-run your tests. If you parse the agent’s tool calls or run tools yourself, you also need to update tool names, parameter names, and how you handle file edits. Either way, the default underlying model changed too, so check output quality and token use before October 5.

What changed in Antigravity Agent 09-2026

The Gemini API release notes for September 17 list the new agent as antigravity-preview-09-2026. The entry says it replaces and deprecates the May version. The May agent launched on May 19, 2026, alongside Google’s announcement of Managed Agents in the Gemini API.

According to the model page for antigravity-preview-09-2026, last updated September 24, the release "upgrades the runtime harness." Google names three parts of that upgrade:

• Improved prompt caching: no figures are given for how much it saves.
• Native code search tools: the agent now has dedicated tools to find files and search their contents.
• Line-range file editing: the agent edits a span of lines instead of rewriting a whole file.

The harness is the layer around the model. It decides which tools the agent can call, how those calls are shaped, and how results come back. Changing it changes what your code sees, even when the model underneath behaves the same.

That is why this matters beyond a version bump. A managed agent is a contract between Google’s runtime and your code. When the tool contract moves, anything that reads tool calls has to move with it.

If you are new to the product, our Google Antigravity 101 covers the wider Antigravity family. Our July piece on the Antigravity SDK and Managed Agents covers the idea of running agentic workflows from your own code.

Who only changes a string, and who changes code

Google’s release note splits users into two groups, and the split is the most useful part of the announcement.

Group one: remote sandbox, final text only. You pass environment: "remote" and read only the output_text or model_output steps. Google says you should "update the agent string and nothing else changes." Most simple integrations fall here. A nonprofit’s content pipeline that asks the agent to research a topic and return a summary is a typical case.

Group two: local tools or parsed tool calls. You run tools yourself through local_environment, or your code reads function_call steps to see what the agent did. For you, the built-in tools changed. Names changed, parameters moved from snake_case to PascalCase, and file edits now arrive as line ranges.

A common group-two pattern is an audit log. Some agencies record every file the agent wrote so a person can review the changes. If that logger looks for a tool called write_file with a path argument, it will find nothing on Antigravity Agent 09-2026. The agent still writes files. Your log just stops recording them.

One detail sits between the two groups. The Antigravity agent guide says filesystem tools such as write_to_file appear as function calls but run automatically in the environment. So even a remote-sandbox user can see these calls in the step list. If you read them for any reason, you are in group two.

The built-in tool names, old and new

Here is the mapping from the September 17 release note. It is the core reference for any Antigravity Agent 09-2026 migration.

Note what did and did not change case. The tool names are still snake_case: write_to_file, view_file, list_dir. The parameters are what moved to PascalCase: TargetFile, StartLine, AbsolutePath. That difference matters for any search-and-replace.

What the new file tools do differently

Four details in the table deserve a second look.

write_file did two jobs in May. It created files and it edited them, always by writing the full content. In Antigravity Agent 09-2026 those jobs split into two tools. Code that treated every write_file call as "file changed" now needs to handle both write_to_file and replace_file_content.

write_to_file adds an Overwrite flag and a Description field. The release note does not define either one beyond the name. Treat them as data to record, not as fields to guess about.

read_file used byte offsets. view_file uses line numbers plus a ContentOffset parameter. Any code that rebuilt file excerpts from byte positions needs to switch to lines.

Search is new. In May, the agent searched by running shell commands such as grep through code_execution. Now find_by_name and grep_search do that work. If you counted shell calls as a proxy for effort, expect those counts to drop.

How line-range edits change your file handling

This is the change most likely to need real engineering time.

Under 05-2026, every edit was a full rewrite. The tool call carried the whole new file. If you wanted the change, you could read content and save it. If you wanted a diff, you compared old and new yourself.

Under Antigravity Agent 09-2026, an edit names a file, a start line, an end line, the content being replaced, and the replacement. That is a patch, not a file. To know the full result, you must apply the patch to the file as it stood before the edit.

Three practical effects follow:

• Diff viewers get easier. The tool call already says which lines changed, so a review screen can show the change directly.
• Mirroring gets harder. If you keep a copy of the agent’s files outside the sandbox, you now have to apply edits in order. Miss one, and every later line number is off.
• Token use should fall on large files. Sending ten changed lines costs less than resending a two-thousand-line file. Google has not published a figure for this, so measure it on your own workloads.

If you do not need the patch detail, there is a simpler route. The environments documentation describes listing and downloading files from a sandbox. For many teams, downloading the finished files at the end is more reliable than rebuilding them from tool calls.

The default model moved to Gemini 3.8 Flash

This part is not in the release note table, and it may affect more teams than the tool changes.

Google’s May 19 announcement described the Antigravity agent as built on Gemini 3.5 Flash. The current guide says that for antigravity-preview-09-2026, the default model is Gemini 3.8 Flash. So a straight string swap also swaps the underlying model, unless you set one.

You can choose the model with agent_config. The guide’s example looks like this:

interaction = client.interactions.create(
    agent="antigravity-preview-09-2026",
    input="Summarize the key differences between functional and object-oriented programming.",
    environment="remote",
    agent_config={
        "type": "antigravity",
        "model": "gemini-3.5-flash-lite",
    },
)

The documented choices are gemini-3.8-flash (the default), gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash, and gemini-3.5-flash-lite. One option during testing is to run Antigravity Agent 09-2026 on gemini-3.5-flash first. That separates the harness change from the model change, so you know which one moved your results.

The guide to building managed agents adds a limit for named agents. For those, the model cannot be overridden at interaction time. You set it in the agent’s definition.

What the Antigravity agent 09-2026 switch does to cost

Google announced no pricing change with this release. The agents overview, updated September 17, describes managed agents as pay-as-you-go on model tokens and tool use. It says one interaction typically uses 100,000 to 3 million tokens. It also says environment compute is not billed during the preview.

The model change is where cost can move. As of September 27, the Gemini API pricing page lists these standard paid-tier rates per million tokens:

On the rate card, the new default is cheaper per token than the May default, now and after the introductory price ends. Per-token rates are only half the bill, though. A model that uses more tokens to finish the same task can cost more at a lower rate. Our look at Gemini 3.8 Flash’s cost per task covers that gap. For agent work, output tokens tend to dominate; our explainer on input versus output tokens covers why.

The practical step is simple. Record token counts per task on 05-2026 now, while it still runs. Then compare the same tasks on Antigravity Agent 09-2026. The guide also documents max_total_tokens inside agent_config as a budget cap, which is worth setting before you move production traffic.

What stayed the same

Several limits did not change in Antigravity Agent 09-2026, based on the two model pages:

• Context window: 1,048,576 input tokens, compacted at about 135,000.
• Output limit: 65,536 tokens.
• Input and output types: text and image in, text out.
• Shell and web search: `code_execution` and `google_search` keep their May names and parameters.
• API surface: both versions are called through the Interactions API with the same `agent` parameter.

The sandbox also looks the same from the outside. Environments idle after 15 minutes, stay resumable for seven days, and are then deleted, per the environments documentation.

What we could not confirm

Some questions a migration plan needs answered are not covered in Google’s documentation as of September 27.

SDK version. The release note does not name a minimum SDK version for Antigravity Agent 09-2026. The Python google-genai package shipped releases on September 16 and September 22, but we could not read their notes to confirm whether either one relates to this agent. Testing on a current SDK release is a reasonable default. It is not a documented requirement.

Local environment details. The release note mentions running tools locally through local_environment. We did not find a documentation page that describes that mode in detail, including how local tool results should be returned under the new parameter names. If you use it, test carefully.

Named agents on the old base. The managed-agents guide, updated September 24, says only antigravity-preview-09-2026 is supported as a base_agent. It also says agent versioning and rollback are not yet available. The docs do not say what happens to an existing named agent built on 05-2026 after October 5. Check yours before that date.

Google Cloud’s Agent Platform. The Managed Agents documentation on Gemini Enterprise Agent Platform refers to the Antigravity harness but, as of its September 25 update, lists no agent versions or shutdown dates. We could not confirm whether the October 5 date applies there.

Antigravity agent 09-2026 migration checklist

Work through these steps before October 5. Each ends with a check you can verify.

1. Find every call site. Search your code, config files, and environment variables for antigravity-preview-05-2026. Check: the search returns zero results after you finish.

2. Change the agent string. Replace it with antigravity-preview-09-2026. Check: a test interaction completes and returns output_text.

3. Decide on the model. Either accept Gemini 3.8 Flash or set model in agent_config while you test. Check: your test log records which model ran.

4. Update tool-name matching. Replace write_file, read_file, and list_files with the new names, and add replace_file_content, find_by_name, and grep_search. Check: a test run that edits a file produces a logged edit.

5. Update parameter names. Change path and content lookups to the PascalCase names in the table above. Check: logged edits show real file paths, not empty values.

6. Rework file-edit handling. Apply line-range edits in order, or download the finished files from the environment instead. Check: your copy of a test file matches the sandbox copy.

7. Replace byte-offset reads. Move any read_file offset logic to the line-based view_file parameters. Check: excerpts shown to reviewers match the file.

8. Compare cost and quality. Run the same tasks on both versions and compare token counts and results. Check: you have numbers for both, recorded before October 5.

9. Set a token budget. Add max_total_tokens in agent_config for production. Check: an oversized test task stops at the cap.

10. Re-run your full test suite. Include approval gates, audit logs, and any alerts built on tool calls. Check: each one fires at least once in testing.

Other October dates to track

If you manage several Gemini dependencies, this is not the only date to track. Our roundup of Gemini model shutoffs in October 2026 collects the others. The deprecation dates are tracked on Google’s deprecations page. We saw a similar pattern in the spring, covered in our piece on the move from Gemini CLI to Antigravity CLI: the rename is easy, and the scripts that depend on output shape take the time.

Frequently Asked Questions