Google Cloud‘s monthly AI recap from January through June 2026 documents a rapid expansion of its agent platform, infrastructure, and model lineup. The centerpiece is the Gemini Enterprise Agent Platform, announced at Next ’26, which provides a unified environment for building, scaling, governing, and optimizing agents. The platform includes a no-code Agent Designer, long-running agents that operate autonomously in cloud sandboxes, an Inbox for managing agent activity, Projects for bounded agent memory, Skills for reusable shortcuts, and Canvas for editing documents. The Agent Gallery offers third-party agents from partners like Adobe, Atlassian, and ServiceNow, with connectors for Asana, Mailchimp, and Workday.

On the infrastructure side, the AI Hypercomputer architecture was updated with two new TPU generations. TPU 8t is optimized for training, using Inter-Chip Interconnect to scale to 9,600 TPUs and 2 PB of shared memory in a single superpod, delivering 3x the processing power of Ironwood and up to 2x better performance per watt. TPU 8i targets inference, connecting 1,152 TPUs via Boardfly topology, with 3x more on-chip SRAM for larger KV caches and a Collectives Acceleration Engine, yielding 80% better performance per dollar than the prior generation. The Agentic Data Cloud was introduced as a new data architecture for agentic AI, featuring a Cross-Cloud Lakehouse standardized on Apache Iceberg, a Knowledge Catalog for dynamic context graphs, and Smart Storage that enriches objects with metadata before agents access them.

Security announcements were extensive. Google Cloud integrated Wiz after completing its acquisition, combining Wiz’s cloud telemetry with Google’s AI and quantum research. New security agents in Google Security Operations help hunt threats, engineer detections, and provide third-party context. Model Armor integrates with Agent Gateway, and Agent Identities provide defenses against shadow AI. Cloud Fraud Defense offers agent-specific capabilities. The Open Knowledge Format (OKF) was introduced as a vendor-neutral standard for representing metadata and curated knowledge, formalizing the LLM-wiki pattern. Google also collaborated with Apple on Private Cloud Compute systems, leveraging Confidential Computing and Titanium security architecture.

Model releases included Gemini 3.5 (Flash and Omni), Gemini 3.1 Pro, Gemini 3.1 Flash Live (scoring 90.8% on ComplexFuncBench Audio), Claude Fable 5, Claude Opus 4.6, Claude Sonnet 4.6, Nano Banana 2 and Nano Banana Pro, Gemini Embedding 2 (first natively multimodal embedding model), Veo 3.1 Lite (at less than 50% the cost of Veo 3.1 Fast), and updates to Veo 3.1 including Ingredients to Video and portrait mode. Early in the year, Google introduced the Universal Commerce Protocol (UCP), an open standard for agentic commerce across discovery, buying, and post-purchase support, compatible with A2A, AP2, and MCP. The Gemini Enterprise for CX platform targets agentic retail. Comments to SQL in BigQuery allows natural language queries. Google also launched the GEAR learning program, published agent production guides, and shared security research from Mandiant and Google Threat Intelligence.