AnswersPublished16 sources
What Google’s Antigravity Preview Adds to Gemini Managed Agents
antigravity preview 09 2026 is Google’s current general purpose Gemini managed agent preview: it defaults to Gemini 3.8 Flash, runs work in a Google hosted Linux sandbox, and adds Files and Credentials APIs. The Files API moves inputs and generated artifacts into and out of an agent environment; the Credentials API...
Published byEdited with GPT-5.6 TerraImages generated with GPT Image 2
Research answer
Key takeaways
- antigravity preview 09 2026 is Google’s current general purpose Gemini managed agent preview: it defaults to Gemini 3.8 Flash, runs work in a Google hosted Linux sandbox, and adds Files and Credentials APIs.
- The Files API moves inputs and generated artifacts into and out of an agent environment; the Credentials API lets an agent connect to services such as GitHub or Slack without exposing the underlying token to the model.
- The preview replaces and deprecates antigravity preview 05 2026; developers can use it through the Interactions API and Google AI Studio, with model choice configurable per interaction.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Google’s release of the antigravity-preview-09-2026 harness for Gemini managed agents introduce, including its native support for G. Article summary: Google’s `antigravity-preview-09-2026` updated the configurable Gemini managed-agent harness, making Gemini 3.8 Flash its default model and bringing its API-hosted tools and behavior closer to Google’s standalone Antigra. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
Google’s September 2026 managed-agent update turns the Antigravity agent into a more capable API-accessible development environment. The new antigravity-preview-09-2026 harness defaults to Gemini 3.8 Flash, brings its tools and behavior closer to Google’s Antigravity Coding Agent, and introduces APIs for file transfer and managed credentials. 5
6
11
What changed in antigravity-preview-09-2026
Google identifies antigravity-preview-09-2026 as the current predefined, general-purpose managed agent. Developers invoke it through the Interactions API, typically with a remote environment that provisions a fresh sandbox for the task. 3
The key model change is Gemini 3.8 Flash as the managed agent’s default. That default is not a lock-in: developers can select the underlying model through agent_config, and can supply instructions, skills, and data inline or save a customized managed-agent configuration for reuse. 1
2
5
Google’s AI Studio announcement says the update brings the managed agent’s tools and behavior closer to the standalone Antigravity Coding Agent. In practical terms, it makes the hosted agent suitable for work that requires more than returning a code snippet: it can work with files, execute code, and pursue a task inside its environment. 4
11
How Gemini managed agents run
A single Interactions API call provisions a Google-hosted secure Linux sandbox, runs the agent loop, and returns the result. Within that environment, the agent can reason about a task, execute code, manage files, and browse the web. 1
3
4
That setup is particularly useful for repository-scale engineering workflows. An agent can inspect a codebase, edit multiple files, run tests or scripts, observe the result, and continue iterating in the same workspace—rather than merely proposing an isolated patch in chat. The available documentation supports filesystem access and code execution, but does not establish specific context-window, maximum-output, or multimodal-input limits for Gemini 3.8 Flash in this release. 1
4
Files API: moving work into and out of the sandbox
The Files API is designed to transfer data between an application and the agent’s environment. Reported functions include uploading inputs to a running sandbox, listing generated files, and downloading individual files or directories. 7
8
10
For a development workflow, that can mean providing a source archive or spreadsheet as input and retrieving edited code, generated reports, or other artifacts after the agent finishes. Google has illustrated the pattern with an uploaded sales spreadsheet that an agent analyzes before producing an interactive dashboard. 15
Credentials API: external access without handing secrets to the model
The Credentials API enables authenticated connections to external services, including examples such as GitHub and Slack, while keeping the actual secret out of the model’s context. 7
9
10
A key security mechanism described in reporting is encrypted credential storage, with the credential inserted only at the outbound connection point rather than exposed as model-readable text. This is a meaningful distinction for agents that need to interact with third-party services: the model can perform an authorized action without being given the reusable token itself. 10
Credentials still require disciplined access design. Keeping a token out of the model context reduces one category of exposure, but teams should continue to scope permissions narrowly and review what actions an agent is allowed to take.
Efficiency and performance claims: what is confirmed
The release is associated with claims of better caching, lower costs, and higher task efficiency. A post from a Google-affiliated engineer, for example, describes up to 22% higher cache-hit rates and lower costs for certain multi-turn coding and reasoning workloads. 14
Those figures should be treated as workload-dependent reported results, not universal benchmarks. The supplied official documentation establishes the model default, managed-agent functionality, and new APIs, but it does not independently document claims such as 40% fewer file-edit output tokens or an 8% increase in multi-turn task completion. Teams evaluating the preview should benchmark it against their own repositories, prompt patterns, tool usage, and latency requirements. 1
3
5
Availability and migration
The updated harness is available through the Gemini Interactions API and Google AI Studio. 4
11
Google’s release notes state that antigravity-preview-09-2026 replaces and deprecates antigravity-preview-05-2026. 6 The supplied official release-note excerpt does not provide a retirement date for the older preview, so an October 5, 2026 cutoff should not be relied on without checking the current migration documentation or account notices.
The practical takeaway
For developers, the important change is not just a model-default swap. Google has combined a configurable Gemini 3.8 Flash agent with a hosted Linux workspace, code and file operations, web access, file ingress and egress, and a credential boundary intended to keep secrets out of model context. That makes antigravity-preview-09-2026 a more complete foundation for controlled, multi-step engineering and automation workflows—provided its performance and security behavior are validated for the specific systems it will touch. 1
2
7
10
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What Google’s Antigravity Preview Adds to Gemini Managed Agents"?
antigravity preview 09 2026 is Google’s current general purpose Gemini managed agent preview: it defaults to Gemini 3.8 Flash, runs work in a Google hosted Linux sandbox, and adds Files and Credentials APIs.
What are the key points to validate first?
antigravity preview 09 2026 is Google’s current general purpose Gemini managed agent preview: it defaults to Gemini 3.8 Flash, runs work in a Google hosted Linux sandbox, and adds Files and Credentials APIs. The Files API moves inputs and generated artifacts into and out of an agent environment; the Credentials API lets an agent connect to services such as GitHub or Slack without exposing the underlying token to the model.
What should I do next in practice?
The preview replaces and deprecates antigravity preview 05 2026; developers can use it through the Interactions API and Google AI Studio, with model choice configurable per interaction.
Sources
ai.google.devWhat's new in Gemini 3.8 Flash | Gemini API | Google AI for Developers
aistudio.google.comGemini API Managed Agents Update - Google AI Studio