New Harness, New Tools: Google's antigravity-preview-09-2026 Rewrites How Its Managed Agents Touch Files
Google's September agent release replaces the May harness: line-range edits instead of full rewrites, PascalCase tool parameters, and native file search — with the old runtime shutting down October 5.
On September 17, 2026, Google quietly reshaped one of the most consequential developer surfaces in its AI stack: the managed agent runtime that powers the Gemini API’s Interactions API and the Antigravity platform. The new agent harness, antigravity-preview-09-2026, replaces and deprecates May’s antigravity-preview-05-2026 — and unlike a routine version bump, it changes the way agents physically interact with code. File edits no longer rewrite whole files. Tool parameters switched from snake_case to PascalCase. The agent now carries its own native file-search tools instead of shelling out to grep and find.
If you build agents for a living, this is the kind of release that looks small in a changelog and looms large in production.
What a Managed Agent Actually Is
First, context. When Google launched managed agents in the Gemini API earlier this year, the pitch was radical simplicity: one API call and Gemini gets a dedicated remote Linux sandbox — a real computer in the cloud that the agent owns for the duration of its task. The model plans, reasons, runs code, manages files, and loops until the work is done, with no infrastructure for the developer to orchestrate. The “harness” is the layer in between: the runtime that defines which tools the model can call, how those tools are parameterized, and how results flow back.
The harness matters more than most people realize. Model quality gets the benchmarks; the harness decides what the model can actually do. And because Google keeps the harness identical between its consumer-facing Antigravity IDE and the API offered to developers, a harness upgrade propagates everywhere at once — from hobbyist projects in AI Studio to enterprise pipelines.
What Changed on September 17
The release notes are unusually precise about who needs to care, and they split users into two camps.
If you run on a remote sandbox (environment: "remote") and only read output_text or model_output steps: nothing to do except update the agent string. The upgrade is effectively invisible.
If you run tools locally (local_environment) or parse function_call steps: the built-in tools changed, and your parsing code will notice.
The differences are structural:
| Capability | 05-2026 | 09-2026 |
|---|---|---|
| File creation | write_file(path, content) | write_to_file(TargetFile, CodeContent, Overwrite, Description) |
| File editing | full rewrite via write_file | replace_file_content(TargetFile, StartLine, EndLine, TargetContent, ReplacementContent) |
| File reading | read_file(path, offset, limit), byte offsets | view_file(AbsolutePath, StartLine, EndLine, ContentOffset) |
| Directory listing | list_files(path) | list_dir(DirectoryPath) |
| File and code search | none — agents used shell commands | find_by_name(SearchDirectory, Pattern, MaxDepth) and grep_search(SearchPath, Query, IsRegex) |
| Shell execution | code_execution(command, timeout_seconds) | unchanged |
| Web search | google_search(queries) | unchanged |
Two shifts stand out.
Line-range editing is the headline. Under the old harness, an agent that needed to change one line in a 500-line file had to rewrite the entire file — an expensive, error-prone pattern familiar to anyone who has watched an LLM “fix” a function by regenerating the whole module and subtly breaking imports elsewhere. The new replace_file_content takes a start line, an end line, the target content, and the replacement. This is how mature engineering tools work: surgical diffs, not wholesale regeneration. It cuts token usage on edit-heavy tasks and, more importantly, makes agent edits auditable — you can see exactly which lines the agent touched.
Native search tools close a long-standing gap. Previously, an agent hunting for a symbol had to drop to the shell and run grep or find as raw commands, parsing free-text output and hoping the command didn’t fail on quoting or permissions. The new grep_search(SearchPath, Query, IsRegex) and find_by_name(SearchDirectory, Pattern, MaxDepth) make filesystem search a first-class, structured tool call. Structured tools mean structured errors, retries, and — critically for anyone evaluating agent frameworks — comparable benchmarks.
The PascalCase parameter names (TargetFile, CodeContent, StartLine) are more than cosmetic. Aligning the API’s tool schema with the conventions used inside the Antigravity IDE tightens the loop between what Google’s own product teams dogfood and what external developers consume. When the internal and external harnesses share a vocabulary, improvements in one surface land on the other without translation.
The Engine Under the Harness
The September harness runs natively on Gemini 3.8 Flash by default — Google’s most capable Flash-class model, generally available since September 2 and engineered, in the company’s words, for “long-horizon software engineering, autonomous agents, and complex enterprise workflows.” Developers can also configure a different default model. The pairing is deliberate: a model tuned for extended, multi-step engineering work benefits most from tools designed for sustained file surgery rather than one-shot generation.
There is also a persistence angle. Google’s own developer advocates describe the upgraded runtime as offering a persistent Linux sandbox — the agent’s environment survives across steps rather than being rebuilt, so state, installed dependencies, and intermediate artifacts carry forward within a task’s lifetime.
The Deadline: October 5
The transition is not optional. antigravity-preview-05-2026 shuts down on October 5, 2026 — tracked on the official deprecations page — leaving an eighteen-day window for anyone parsing function_call steps or running tools locally to update their integration. Teams that only consume final outputs can switch the agent string and move on. Teams with custom tool-call parsers, logging pipelines, or evaluation harnesses built around snake_case parameters and full-file writes have real migration work: every parser, mock, and fixture keyed to the old schema needs revisiting.
It is a pattern the industry is still learning: in the agent era, the deprecation cycle applies not just to models but to the tool contracts around them. Your code may not call write_file directly — but if you parse the calls your agent makes, a schema change is an API break all the same.
Why It Matters
The agent-harness layer is becoming the real battleground of agentic AI. OpenAI’s Codex, Anthropic’s Claude Code, and Google’s Antigravity all increasingly compete not on model benchmarks alone but on the quality of the runtime that lets a model act on real repositories: how it reads, how it edits, how it searches, how it recovers from mistakes. Google’s September release reads as a direct answer to that competition — borrowing the proven idioms of professional developer tooling (diff-based editing, structured search, explicit parameters) and pushing them into the default experience of every managed agent built on the Gemini API.
For developers, the takeaway is straightforward. If you touched function_call parsing or local tool execution, schedule the migration before October 5. If you haven’t looked at managed agents yet, the pitch just got stronger: a single API call now buys you a persistent Linux sandbox, a coding-tuned frontier model, and a tool vocabulary that finally speaks fluent IDE.
The unglamorous parts of agent infrastructure — file editing conventions, search primitives, parameter schemas — are where agent reliability is actually won. Google’s September harness upgrade is a sign the company knows it.