Google has officially upgraded its Gemini Managed Agents framework with a fresh backend release: the antigravity-preview-09-2026 harness. Powered under the hood by Gemini 3.8 Flash, this upgrade addresses two of the biggest hurdles developers face when building autonomous AI workflows—latency on extended tasks and secure integration with external enterprise tools.

The release introduces intelligent model switching, significant context caching gains, and two dedicated APIs designed to simplify data ingestion and authorization.

The new harness acts as the execution backbone for Gemini Managed Agents, orchestrating how the underlying model handles long-running multi-step tasks, system instructions, and dynamic tool calls.

By leveraging Gemini 3.8 Flash, the platform delivers faster token throughput without sacrificing reasoning quality.

• Optimized Context Caching: Google overhauled how state is preserved across agent loops. Context caching efficiency has increased significantly, resulting in a 22% hit rate improvement on long, multi-step tasks and a 9% boost on complex coding workflows. This reduction in redundant token processing translates directly to lower latency and reduced API costs.

• Dynamic Model Switching: Agents can now seamlessly pivot between task-specific model variants depending on the execution step. For instance, an agent can use a lightweight mode for basic text extraction, then switch to high-reasoning execution for data analysis or code generation.

• Streamlined Agent Loops: Redundant state serialization between steps has been minimized, preventing “agent drift” during long-running background tasks.

To make AI agents genuinely useful in production environments, they need direct access to unstructured data and external authentication systems. Google is bridging this gap with two API additions.

The Files API allows developers to stream large datasets, media files, and unstructured documents directly into an agent’s working context.

Instead of writing custom pre-processing pipelines or complex vector indexing steps, you can pass structured files—such as sales spreadsheets, financial reports, or codebases—straight to the agent.

Real-World Example: An agent can consume a raw Q3 sales spreadsheet via the Files API and autonomously generate a real-time, interactive performance dashboard displaying metrics like $2.67 million in Q3 revenue, categorized by region and product tier.

One of the primary risks when deploying autonomous agents is the exposure of API keys, OAuth tokens, and database secrets.

The new Credentials API acts as a secure vault for external integrations. It allows Gemini Managed Agents to call third-party services—such as Jira, Salesforce, GitHub, or internal databases—without exposing raw credentials to the model context or logging pipelines. Secrets remain encrypted at rest and in transit, executing strictly through scoped, ephemeral access tokens.

Key figures across Google’s developer relations and engineering teams highlighted how user feedback drove these specific performance and security updates.

• Philipp Schmid noted the importance of the caching upgrades, emphasizing that a 22% improvement in context cache hits fundamentally alters the unit economics of building long-context agents.

• Logan Kilpatrick pointed to the developer experience, highlighting how combining file handling with credential management turns Gemini into a far more practical environment for production-grade builds.

The antigravity-preview-09-2026 harness, along with the Files and Credentials APIs, is available immediately across Google’s developer surface areas:

• Google AI Studio: Developers can prototype agents using a visual interface, configure system prompts, attach datasets via the Files API, and test tool integrations directly.

• Interactions API: Production applications can leverage these features programmatically with updated SDK support.

• Free Tier Access: Google has included access to these new harness capabilities within the standard free tier limits in Google AI Studio, allowing teams to test performance prior to scaling.

For more real-time leaks, mobile updates, and AI coverage, head over to NokiaPowerUser.

Add NPowerUser as a preferred source on Google News