Google Adds Offline AI Agent Workflows to Antigravity SDK

An Antigravity agent can now work through a coding task without sending its model requests to a cloud service. Google added local-model support to the developer SDK, starting with Gemma 4 26B A4B running through LiteRT. The choice gives developers a way to keep code and requests on their machine, though Google recommends more than 24GB of graphics or unified memory for its example.

The agent stays; the model moves

The Antigravity SDK lets developers build agents that use a model to carry out multi-step work. With the update, that model can run locally while the SDK still handles the surrounding agent workflow. LiteRT is Google’s on-device runtime for the initial Gemma integration: developers download the model, point the SDK’s configuration at its file and run the agent against it.

Developers need not use that pairing. The SDK can connect through LocalOpenAIAgentConfig to an OpenAI-compatible local inference server, including Ollama, LM Studio or vLLM. Google says developers can switch among those backends without changing their agent orchestration, tools or workflows. That makes the update more than a route to one Gemma model: it offers a way to retain an existing local model server while using Antigravity’s agent framework.

The machine takes on the work

For a fully local workflow, Google points to three advantages: no model API charges or rate limits, the ability to keep code and requests on the machine, and operation without a stable connection once the setup is in place. The tradeoff is that the developer supplies the compute. Google recommends a machine with more than 24GB of VRAM or unified memory for its Gemma example, and its sample warns that local inference could take several minutes.

Google’s smaller demonstration gives a concrete picture of local execution. From one prompt, a Gemma-powered agent writes a Python terminal monitor for CPU and memory use, creates a requirements file and tests the script. The work runs on the developer’s machine. It shows the sort of bounded task Google chose to demonstrate, not how the setup performs across other coding jobs.

A cloud planner, with a narrower view

The hybrid option makes a different bargain. In Google’s recorded run, Gemini 3.8 Flash assigned work for an audit of three vulnerable Python modules. Local Gemma agents then reproduced vulnerabilities, proposed fixes, critiqued patches and ran regression tests. Google says the run used 95 cloud tokens and 3,322 local tokens, putting 97.2% of token use on-device. It says the resulting patches passed the tests in that demonstration.

That division protected the source files from the cloud model in Google’s example, but it did not make the workflow wholly private or offline: filenames and task descriptions still went to Gemini. A team deciding whether to use the hybrid pattern would need to decide if those details can leave its machine. The recorded run also cannot establish how reliably local agents will find and repair bugs in other projects.

Google’s published examples also show that choosing a local model is only one part of setting up an agent that changes files. The resource-monitor sample gives the agent a working directory and sets its policy to allow all operations. Developers adapting that example must make their own decisions about what an agent may do in its workspace. Local inference determines where the model runs; it does not, by itself, decide which actions the agent should be permitted to take.