Google Cloud Launches Agent Compute Engine for Scalable AI Agents

On October 22, 2026, Google Cloud unveiled its new Agent Compute Engine at the AGNTCon + MCPCon events. This compute service is specifically designed to operate AI agents at scale, utilizing hardware-isolated microVMs for enhanced security and efficiency.

The Agent Compute Engine allows for rapid deployment of AI agents, enabling them to start quickly, pause, and resume operations as needed. A key feature is the copy-on-write memory sharing, which allows multiple agents to share memory until changes are made, significantly increasing the density of agents that can be run on the same hardware.

Google emphasizes the importance of security in this new infrastructure, leveraging its Titanium IPU technology to protect against potential threats. The service is part of the broader Gemini Enterprise Agent Platform, which focuses on security and governance for AI agents.

Currently, the Agent Compute Engine is already in use internally for Google’s Gemini agents, with a preview phase for external customers expected to begin soon. This development aims to address the challenges of scale and containment that arise from the dynamic nature of AI agents.

New Developments in AI Text Detection

• Research Collaboration: Google DeepMind and the University of Pennsylvania have developed a new method for AI text detection that enhances transparency.
• Publication Details: The research paper titled "Building Auditable AI Text Detectors with Sparse Autoencoders" was submitted on October 21, 2026, and accepted at EMNLP 2026.
• Performance Metrics: The interpretable classifier achieved an F1 score of 96.8 on in-domain data, compared to the black-box detector's score of 98.0. On out-of-distribution data, the new model scored 92.2, while the black-box dropped to 75.2.
• Feature Explanation: The new method utilizes sparse autoencoders to convert complex internal signals into human-readable features, enhancing the auditability of AI decisions.
• Public Availability: The code and trained models are available on GitHub under the google-deepmind/detectorscope repository.