Google Antigravity SDK Adds Local Model Support for Hybrid Agent Workflows

Google’s Antigravity SDK supports hybrid agent workflows that combine local model execution with cloud-based models, enhancing data privacy, cost control, and workload scheduling. At its core is the implementation of the “Architect-Builder” pattern within a hybrid environment.

Key Insights:

• The Architect (Cloud): Decomposes plans based solely on task descriptions and filenames, without needing access to source code.
• The Builder (Local): Focuses on executing code-intensive tasks such as vulnerability reproduction, patch review, and regression testing.
• Hybrid Orchestration: Achieves separation of duties by keeping sensitive computations local while offloading high-level reasoning to the cloud.

This article is suitable for AI system architects, security tool developers, and MLOps engineers focused on privacy compliance.

Google has introduced support for local models in the Antigravity SDK, enabling developers to build and execute agentic workflows directly on their own machines while retaining the flexibility to integrate cloud models for higher-level tasks. This allows developers to better manage data privacy, inference costs, network dependencies, and the distribution of computational workloads.

One of the most compelling features of the Antigravity SDK is hybrid agent orchestration. This feature applies Google’s proposed “Architect-Builder” pattern to a hybrid environment composed of both cloud and local computing resources.

According to the “Architect-Builder” pattern, a primary AI agent with high contextual awareness (the “Architect”) maintains a global view, including dependencies, architectural constraints, and detailed specifications or implementation plans. Once the Architect formulates a plan, multiple parallel AI agents with specialized expertise (the “Builders”) operate independently in isolated environments, executing narrow, discrete tasks outlined in the specification.

Google states that the Antigravity SDK enables the “Architect” to run in the cloud, formulating strategies and breaking down tasks into smaller units using only filenames and task descriptions, without accessing the source code itself. Meanwhile, the “Builders” run locally, performing tasks such as reproducing security vulnerabilities, proposing code fixes, reviewing candidate patches, and running regression tests. Other specialized local agents can validate proposed changes against provided test suites, thereby keeping code-intensive stages of the workflow on-device.

In a demonstration involving auditing and patching three vulnerable Python modules, Google reported that out of approximately 3,400 tokens, only 95 were used for cloud planning, while the remaining 97.2% of tokens were consumed locally. Google also confirmed that no source code was uploaded to the cloud.

The initial implementation primarily focuses on running the Gemma 4 26B A4B model via Google’s LiteRT technology. However, developers are not limited to LiteRT, as the Antigravity SDK also supports OpenAI-compatible local inference servers—including Ollama, LM Studio, and vLLM—through LocalOpenAIAgentConfig.

Regarding this announcement, Reddit user ambassadortim noted: “Enterprises can have more control over their AI models, including when changes are made.” In another discussion, user FollowingTop3534 questioned whether the metric of cloud token share fully reflects the situation:

I would benchmark patch acceptance rates, runtime, and energy consumption alongside cloud token share. While the statistic that “97.2% of token consumption came from local models” is interesting, it doesn’t necessarily indicate whether running models on your hardware is faster or cheaper. Furthermore, recommending 24GB+ of VRAM or unified memory does present a barrier to entry.

As a final note on community discussions, user Latter_Crazy posted an enthusiastic comment about these new capabilities:

It achieves things I couldn’t do previously with cloud rules. Costs, speed, and quality are all improving, and this is just the beginning. Highly recommended for everyone to check out.

Please do not confuse the Antigravity SDK with the Antigravity IDE, which is an Integrated Development Environment (IDE) powered by the Antigravity SDK. They share the same underlying agent infrastructure but serve different purposes: the SDK allows you to programmatically build your own AI agents, extending beyond just coding tasks.