What changed
Google released antigravity-preview-09-2026 on September 17 as the replacement for its first Antigravity Agent preview. The model is exposed through the Gemini Interactions API: one request can provision a Google-hosted Linux sandbox, run the agent loop, and return the result. Google describes the agent as a general-purpose runtime that can reason, execute Bash/Python/Node commands, manage files, and browse the web.
The release is more than a new model identifier. The managed agent uses the same harness as the Antigravity IDE, defaults to Gemini 3.8 Flash, and lets builders override the underlying model with agent_config. The API also keeps the sandbox environment separate from conversation state, so a follow-up can preserve files while starting a new conversation—or keep the conversation while creating a fresh workspace.
Why it matters
This is a shift from calling a model to renting an agent runtime. A product team can now prototype a research worker, coding assistant, or document-processing workflow without assembling its own shell runner, file workspace, browser layer, and session persistence first. The Interactions API supplies the surrounding primitives too: observable execution steps, server-side conversation state, background execution for long-running tasks, and an explicit storage control.
The practical tradeoff is migration discipline. Local-tool calls in the new preview use PascalCase parameters, and file edits move from full-file rewrites to line-range replacements. The old antigravity-preview-05-2026 preview is deprecated and scheduled to shut down on October 5, so integrations that inspect function-call steps or execute local tools should be updated before switching versions.
Builder takeaway
Treat the agent version and sandbox as part of your product contract. Pin the preview identifier, keep tool execution behind an approval boundary, persist only the files your workflow needs, and test both the agent’s final answer and its intermediate steps. That makes it easier to upgrade the underlying model without rewriting the application’s orchestration layer.