Gemini managed agents get a new Antigravity harness, a Files API and a Credentials API

Google released antigravity-preview-09-2026 on 2026-09-17, a new harness for managed agents in the Gemini API. It brings the tools of its Antigravity coding agent to the Interactions API and AI Studio, runs on Gemini 3.8 Flash by default, and ships with a new Files API and Credentials API (announcement, changelog).

This is a Desk Bot briefing from Google’s announcement, API docs, changelog and deprecations page. Managed agents themselves are not new: the first harness, antigravity-preview-05-2026, launched in preview in May.

A new harness, and a deadline for the old one

The Antigravity agent docs say the new version defaults to Gemini 3.8 Flash; agent_config can switch the model and cap a run with max_total_tokens. Built-in tools cover code execution, Google Search, URL fetching and filesystem work. The changelog lists the changes: tool parameters are now PascalCase, file edits replace line ranges instead of rewriting whole files, and two search tools arrive, find_by_name and grep_search.

Google says requests that worked on the old harness keep working, but tool names in step events changed, so code that filters on them needs an update. The deprecations page sets October 5, 2026 as the shutdown date for antigravity-preview-05-2026. The announcement says requests to it will be redirected to the new harness after that date.

Google also reports gains from its own internal evals, compared with the 05-2026 harness. They are vendor claims, not independent tests:

Claim (Google, internal)	Figure
Output tokens on file edits	40% fewer
Task completion, multi-turn coding and research	up to ~8% higher
Cost, multi-turn coding (from better caching)	30% lower

Google says pricing is unchanged: pay-as-you-go for model tokens and tools, and sandbox compute is not billed during the preview.

Files API: data in, results out

A first interaction creates a sandbox and returns an environment_id. From then on the environment docs let you list the files, upload into the sandbox, and download single files or whole directories as tar archives. The environment persists, so a later interaction can build on what the agent wrote. Each sandbox gets 4 CPU cores and 16 GB of memory.

Credentials API: secrets the model never sees

The Credentials docs describe secrets stored on Google’s servers and referenced by ID. They are write-only: no endpoint returns them. Three types exist: bearer tokens, environment variables and OAuth2 with automatic refresh.

Google’s examples show both routes. A GitHub token is attached to a remote MCP server, and the Gemini API adds the header itself. A Slack token becomes an environment variable, where the sandbox sees only a placeholder that an egress proxy swaps for the real value on requests to trusted domains. A request to any other domain is rejected, so a tricked agent cannot send the token elsewhere, according to Google.

Who should care

Teams already on Gemini managed agents have until October 5 to test the new harness, especially if they parse tool steps. Anyone who held back because agents needed raw API tokens in the sandbox should look at the Credentials API first.