Gemini Managed Agents Harness Update Adds Files API and Credentials API

Gemini managed agents are a general-purpose managed agent on the Gemini API that reasons, executes code, manages files, and browses the web inside a secure Linux sandbox.

The antigravity-preview-09-2026 harness became available in September 2026 and now serves as the default for Gemini managed agents in both the Interactions API and AI Studio.

Existing requests that run on the prior harness continue without interruption, and pricing for the service stays the same.

What background led to the antigravity-preview-09-2026 harness release?

The earlier antigravity-preview-05-2026 harness supported core agent functions including reasoning and code execution inside a secure Linux sandbox.

Developers using the previous version encountered constraints when handling persistent data or external service authentication, often requiring custom workarounds that reduced efficiency.

Google designed the new harness to align agent behavior more closely with capabilities available in the Antigravity Coding Agent while adding dedicated APIs for file and credential operations.

What new features does the updated harness include?

The Files API permits uploading data directly into active agent sandboxes, listing available files, and retrieving specific files or entire directories from the sandbox environment.

The Credentials API supports one-time registration of secrets in the form of bearer tokens, environment variables, or OAuth2 credentials that attach to MCP servers or sandbox environment variables.

Improved caching mechanisms deliver higher hit rates of approximately 9 percent on multi-turn coding and research tasks and 22 percent on long question-answering runs.

1. Register credentials through the Credentials API using one of the supported secret types.
2. Upload required data files into the sandbox via the Files API before agent execution begins.
3. Configure the agent instance to default to Gemini 3.8 Flash for the interaction.
4. Execute the agent task and retrieve outputs including any downloaded files or directory contents.

How do the technical specifics of the new APIs operate?

The Files API operates within the secure Linux sandbox environment, enabling the agent to perform data management operations without exiting the controlled runtime.

The Credentials API relies on a network proxy positioned around the agent sandbox to inject registered credentials during outbound requests.

We created a network proxy around the agent sandbox which basically injects the credentials when the agent makes a request from inside the sandbox to outside the sandbox. So the agent never really sees your credential.Philipp Schmid, Member of the Technical Staff, Google DeepMind

This proxy approach ensures that the model itself never receives or processes the actual credential values during operation.

Comparison of features between the previous and updated Gemini managed agent harness versions

Aspect	antigravity-preview-05-2026	antigravity-preview-09-2026
Default Model	Gemini 3.5 Flash	Gemini 3.8 Flash (configurable per interaction)
File Management	Limited capabilities	Full Files API for upload, list, and download
Credential Handling	Manual workarounds required	Credentials API with network proxy injection
Cache Hit Rate Improvement	Baseline	Approximately 9% on coding tasks, 22% on long QA
Deprecation Status	Active until October 5, 2026	Current default with automatic redirection after deprecation

What cost and performance changes result from the harness update?

The same caching gains produced a 17% cost reduction on reasoning tasks according to measurements reported by Google AI Studio.

These efficiency improvements apply to tasks that involve repeated interactions within the agent sandbox without altering the underlying model pricing structure.

What market and stakeholder implications follow from the update?

Developers focused on multi-turn coding and research workflows gain direct access to lower operational costs through the elevated cache performance.

Enterprise users benefit from the secure credential registration process that reduces the need for custom security layers around agent deployments.

The shift to Gemini 3.8 Flash as the default model provides access to an updated base capability while preserving the option to select alternative models per interaction.

What expert reactions address the harness changes?

Philipp Schmid, Member of the Technical Staff at Google DeepMind, described the credential handling mechanism in detail.

What comes next for Gemini managed agents?

The antigravity-preview-05-2026 harness will reach deprecation on October 5, 2026, after which all requests will redirect automatically to the new harness.

Developers are advised to test existing agent configurations against the updated APIs to ensure compatibility ahead of the transition date.

Subsequent releases may extend the Files API and Credentials API with additional integration points for other Google services.