Google Ships antigravity-preview-09-2026 Harness for Gemini Agents With Files API and Credentials API
The update that teams building on the Gemini Interactions API have been waiting for: 40 percent fewer output tokens on file-edit operations, a secure path for agents to call external services without credentials in the prompt, and a native way to move files in and out of the agent sandbox. Google released all three together in the antigravity-preview-09-2026 harness on September 17, 2026.
What shipped
The new harness runs on Gemini 3.8 Flash with the model configurable per interaction, per the Google AI Studio announcement. On file editing, it produces 40 percent fewer output tokens and delivers up to 8 percent higher task completion on multi-turn software engineering and research benchmarks. Cache hit rates improved by roughly 9 percent on multi-turn coding and research tasks and by 22 percent on long question-answering runs. Per Google, those cache gains translate to 17 percent lower cost on reasoning tasks and 30 percent lower cost on multi-turn coding runs.
Two new APIs ship alongside the harness. The Files API lets developers upload data directly into the agent's isolated Linux sandbox before a session and download individual files or directories when it finishes. The Credentials API lets the agent call services the developer already has access to, from remote MCP servers to third-party REST APIs, without storing API keys in the prompt or system instruction. Per the announcement, secrets are injected at runtime and the model never handles the raw credential values.
Tool names in step events changed in the new harness: file tools now report as view_file, write_to_file, and replace_file_content rather than the previous names. Any code filtering on step event tool names requires a corresponding update.
The prior harness, antigravity-preview-05-2026, will be deprecated on October 5, 2026. Per the Gemini API docs, existing requests redirect automatically to the new version after that date. Pricing is unchanged from the prior harness. The update is live in AI Studio and the Gemini Interactions API now.
Agents can reach internal APIs without credentials in prompts
The Credentials API removes the main security objection that slows enterprise adoption of managed agents. Before this release, giving a Gemini agent access to an internal MCP server or an OAuth-gated REST API meant passing credentials through the prompt or building a separate proxy layer. Teams can now scope access at the interaction level and rotate credentials outside the model context entirely.
The 40 percent reduction in output tokens on file edits compounds on agentic coding jobs that involve many sequential small file changes. Combined with the 30 percent cost drop on multi-turn coding runs, the economics of running software engineering tasks through the Interactions API shift meaningfully. For teams already on the API, the numbers argue for re-running cost baselines before the October cutover.
What to watch next
The October 5 deprecation window gives teams roughly two weeks to verify that the auto-redirect does not break step-event filtering or any logic tied to the previous tool names. Whether Google publishes a code-sample migration guide before the cutoff is the immediate question. The harness also ships with configurable model selection per interaction, so how operators mix in lighter Gemini variants for cost-constrained tasks is a usage pattern worth tracking as the harness matures.
Sources
- Gemini API Managed Agents Update: Google AI Studio, September 2026
- Antigravity agent documentation: Google AI for Developers, September 2026
- Google Upgrades Gemini Managed Agents: Nokia Power User, September 18, 2026
