Story

Google Boosts Gemini AI Agents with Efficiency Upgrades and New APIs

ENTHMSVIIDZHZH-TWJAKOHI
Sep 17, 20262 min read
Google Boosts Gemini AI Agents with Efficiency Upgrades and New APIs

Summary

Google has released an update for its Gemini managed agents, introducing new file management and security APIs while delivering significant performance improvements, including a 40% reduction in token usage for certain tasks.

Text size
Background

Google announced on Thursday an updated harness for its Gemini managed agents, a move aimed at enhancing performance, security, and functionality for developers using its artificial intelligence platform.

The new release, identified as `antigravity-preview-09-2026`, runs on the company's Gemini 3.8 Flash model and is designed to make the AI agents more efficient and capable for complex tasks.

Key Performance Gains

According to internal testing conducted by Google, the updated agent harness delivers notable efficiency improvements. These upgrades can translate into lower operational costs and faster results for developers building on the platform.

Key metrics reported by the company include:

  • A 40% reduction in the use of output tokens when making file changes.
  • An increase of up to 6% in task completion rates for multi-turn software engineering and research assignments.
  • An up to 16% improvement in cache hit rates during long conversations and research tasks.
Sample IUX Markets – In-articleAd

New Developer Tools and Security

The update introduces two significant new Application Programming Interfaces (APIs) designed to expand the agents' capabilities within their secure environment. Gemini managed agents provide developers with an isolated Linux sandbox where the AI can execute code, browse the web, and modify files to complete objectives.

  • The Files API allows users to directly upload data into the agent's sandbox, list the files it creates, and download specific files or entire directories.
  • The Credentials API enhances security by enabling the agent to connect to external services like GitHub and Slack without directly exposing sensitive access tokens to the AI model. Credentials are encrypted and only applied at the final point of connection.

Availability and Transition

The new harness is now available for developers in Google AI Studio and through the Interactions API, with pricing remaining unchanged from the previous version. Google stated that the older harness, `antigravity-preview-05-2026`, will be deprecated on October 5, 2026. After this date, all requests will be automatically redirected to the updated version.

Read next

More on Stocks
Back to latest news

LATEST