Google adds local model support to Antigravity SDK

On 2026-10-29, InfoQ reported that Google has added local model support to Antigravity SDK. Google said developers can now build and run agentic workflows directly on their own machines while still folding in cloud models for higher-level tasks, giving them more control over data privacy, inference cost, network dependency and how compute work is split. The company described the headline capability as hybrid agent orchestration, which applies its Architect-Builder pattern to an environment that mixes cloud compute with local compute. In that pattern, a high-context-awareness lead AI agent, the Architect, holds the full picture, including dependencies, architecture constraints and detailed specs or implementation plans. Once the Architect has a plan, multiple parallel specialized AI agents, the Builders, work independently in isolated environments on narrowly scoped discrete tasks. Google said Antigravity SDK lets the Architect run in the cloud, forming strategy and breaking tasks into smaller units from file names and task descriptions alone, without needing access to source code itself. The Builders run locally and take on work 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, keeping the code-heavy stages of the workflow on the device. Google also gave a vulnerability audit demo covering three vulnerable Python modules, stating that of roughly 3400 tokens, only 95 went to cloud planning, and the source code was never uploaded to the cloud. The report added that the SDK works with local inference servers including Ollama, LM Studio and vLLM. So far Google has not published a separate version number for this capability, and no further developments have been reported.