Google Cloud CEO Thomas Kurian opened Cloud Next 2026 in Las Vegas on 22 April with a keynote titled 'The Agentic Cloud,' delivering the company's most consequential infrastructure announcement since the original TPU launch. The headline hardware reveal was Ironwood, Google's seventh-generation Tensor Processing Unit, which delivers 4.6 petaFLOPS per chip and scales to 42.5 exaFLOPS in 9,216-chip superpods. On the model side, Google announced Gemini 3 Pro and Flash — with Flash delivering a 15 per cent accuracy improvement over Gemini 2.5 Flash — alongside Gemini 3.1 Pro in preview and an experimental model called GLM 5 targeting complex systems engineering. Vertex AI was formally rebranded to Gemini Enterprise Agent Platform, absorbing Google Agentspace into a unified product.
The most architecturally significant announcement for developers was the Agent-to-Agent (A2A) protocol reaching version 1.2 with cryptographic signatures and 150 organisations now routing real tasks between agents in production. Microsoft, AWS, Salesforce, SAP, and ServiceNow are all running A2A in their environments, making it the first cross-platform agent communication standard to achieve meaningful enterprise adoption. Google also announced that Apigee now functions as an MCP bridge, translating existing APIs into discoverable agent tools — effectively connecting Google's A2A protocol with Anthropic's Model Context Protocol ecosystem. Fully managed MCP servers for Google Maps, BigQuery, Compute Engine, and Kubernetes Engine ship alongside the announcement.
For the developer toolchain, Google released Agent Development Kit v1.0 with stable libraries across Python, Go, Java, and TypeScript, a visual Agent Designer canvas in preview, and Agent Engine Sessions with Memory Bank for persistent context across interactions. Workspace Studio brings no-code agent building to business users, enabling automations across Gmail, Docs, Sheets, and third-party apps like Jira, Salesforce, and Asana using plain language prompts. Model Garden now hosts over 200 models including Anthropic Claude and open models like Llama. For context engineers, the convergence of A2A and MCP through the Apigee bridge is the detail that matters most — it signals that the two competing agent communication standards are not going to war but are instead being bridged at the infrastructure layer, which means developers can build on either protocol without being locked into a single ecosystem.
