The developer-facing platform, now called the Gemini Enterprise Agent Platform, received deeper upgrades. Agent Designer, a visual flow canvas for building agent workflows, is in preview. Agent Engine Sessions and Memory Bank, which give agents persistent context across interactions, are generally available. A new Agent Garden provides prebuilt agent solutions for customer service, data analysis, and creative tasks. A free tier via Express mode lowers the entry barrier. The Model Garden now hosts more than 200 models spanning Google's own Gemini and Gemma families, third-party models including Anthropic Claude, and open models such as Llama. Google also announced six new agents for data engineering and coding in BigQuery, including a data engineering agent that automates pipeline creation from natural language prompts and a code interpreter that translates queries into executable Python with visualisations. Partner agents from Box, Workday, Salesforce, ServiceNow, Dun and Bradstreet, and S&P Global are integrated into the platform, giving enterprise customers prebuilt capabilities for document intelligence, HR self-service, IT operations, and financial data.

Project Mariner, Google DeepMind's web-browsing agent powered by Gemini 2.0, scores 83.5% on the WebVoyager benchmark and handles ten concurrent tasks on cloud-based virtual machines. It automates shopping, information retrieval, and form-filling, and is available to Google AI Ultra subscribers in the United States. The roadmap includes a visual builder called Mariner Studio in the second quarter, cross-device synchronisation in the third quarter, and an agent marketplace in the fourth quarter.

The protocol play

The most strategically significant announcement may be the least visible to end users. Google's Agent2Agent (A2A) protocol, originally launched with more than 50 technology partners, has reached 150 organisations in production, not pilot, routing real tasks between agents built on different platforms. The protocol is now governed by the Linux Foundation's Agentic AI Foundation and has reached version 1.2, with signed agent cards using cryptographic signatures for domain verification. Microsoft, AWS, Salesforce, SAP, and ServiceNow are running A2A in production environments.

A2A is designed to complement rather than compete with Anthropic's Model Context Protocol (MCP). MCP handles how an agent connects to tools and data sources. A2A handles how agents communicate with each other across organisational and platform boundaries. Google adopted MCP across its own services in December 2025, launching fully managed remote MCP servers for Google Maps, BigQuery, Compute Engine, and Kubernetes Engine, with Cloud Run, Cloud Storage, AlloyDB, Cloud SQL, Spanner, Looker, and Pub/Sub on the roadmap. Apigee, Google's API management platform, now functions as an MCP bridge, translating any standard API into a discoverable agent tool with existing security and governance controls. Google is simultaneously positioning A2A as the standard for the layer above: the orchestration of multiple agents from multiple vendors working together on a single task.

The practical implication is that a Salesforce agent built on Agentforce can hand off a task to a Google agent running on Vertex AI, which can query a ServiceNow agent for IT asset data, all through A2A without any of the three systems needing to understand each other's internal architecture. Native A2A support is now built into Google's Agent Development Kit, LangGraph, CrewAI, LlamaIndex Agents, Semantic Kernel, and AutoGen. Google's open-source Agent Development Kit reached stable v1.0 releases across Python, Go, and Java, with TypeScript support also available. It is a code-first framework optimised for Gemini but model-agnostic and deployable to any container or Kubernetes environment. The security layer includes Model Armor for defence against indirect prompt injection, zero-trust architecture applied to decentralised agent systems, and access management through Google Cloud IAM with audit logging.

The competitive landscape

OpenAI's own enterprise agent push through Codex and systems integrator partnerships has reached three million weekly users. Anthropic's enterprise marketplace for Claude-powered tools is building an ecosystem through partners including Snowflake. Microsoft's Copilot is embedded in virtually every Fortune 500 company. AWS has Bedrock with its own agents framework maturing rapidly. The enterprise AI agent market is not a two-horse race. It is a five-way contest in which each competitor has a structural advantage the others lack.

OpenAI has the strongest consumer brand and the most advanced reasoning models. Anthropic has the most trusted safety positioning and the fastest-growing enterprise revenue. Microsoft has the deepest enterprise distribution through Office and Azure. AWS has the largest cloud infrastructure base and the strongest developer gravity. Google's argument is that it is the only company that owns all four layers of the stack: the custom silicon (Ironwood TPUs), the frontier models (Gemini), the cloud platform (now unified as the Gemini Enterprise Agent Platform), and the enterprise distribution channel (Workspace with more than three billion users across Google's productivity tools). Kurian framed the strategy explicitly: “If you want to adopt a technology successfully, you need to pick a few important projects and do them well, rather than spraying on a lot of little projects.” No other competitor controls the full vertical from chip to application.

Google's own AI Agent Trends report, published ahead of the conference, found that 89% of business teams are already using AI agents and the average organisation runs 12. The most common enterprise use cases are customer service at 49%, marketing at 46%, security operations at 46%, and IT support at 45%. Early customer deployments suggest the productivity claims are not purely theoretical: Danfoss, the Danish industrial manufacturer, automated 80% of transactional decisions in email-based order processing using Google's agents, reducing response times from 42 hours to near real-time. Suzano, a Brazilian pulp and paper company, built an agent with Gemini Pro that translates natural language into SQL queries, cutting query time by 95% for 50,000 employees.

The model foundation

The agents run on Google's Gemini model family, with the Gemini 2.5 generation being retired in October in favour of the 3.x line. Gemini 3 Pro and Gemini 3 Flash, released in late 2025 and iterated through early 2026, provide the reasoning backbone. Gemini 3 Flash delivers a 15% improvement in overall accuracy over Gemini 2.5 Flash and is optimised for high-frequency agentic workflows and real-time processing. Gemini 3.1 Pro, the most advanced reasoning variant, is available in preview. A new experimental model, GLM 5, targets complex systems engineering and long-horizon agentic tasks through the Model Garden. Gemini 3.2 is expected to be formally announced during the conference, with an expanded context window beyond one million tokens and optimised parameter counts for reduced inference latency. Demis Hassabis, DeepMind's chief executive, stated in January that his team is “focusing on Gemini 4 this year.” Google also recently launched Gemma 4 open models under Apache 2.0 licensing, built from the same research as Gemini 3 and providing an open-weight alternative for enterprise customers who need to run models on their own infrastructure.

The infrastructure beneath the models is equally central to the pitch. Ironwood, Google's seventh-generation TPU announced the same day, delivers 4.6 petaFLOPS per chip and scales to 9,216-chip superpods producing 42.5 exaFLOPS. Anthropic has committed to up to one million Ironwood units. The custom silicon means Google can offer inference at costs that customers buying Nvidia GPUs at retail cannot match, which in a market where inference is the dominant and growing expense, translates directly into pricing power for the agent services that run on top.

Google Cloud holds roughly 11% of the cloud infrastructure market. AWS holds 31%. Azure holds 25%. The gap is significant and Cloud Next will not close it. But the agentic era, if it materialises at the scale Google is projecting, reshuffles the competitive dynamics in ways that favour a company with a vertically integrated stack over companies that assemble their AI capabilities from multiple vendors. Google is betting that the enterprise customer who adopts AI agents at scale will choose the platform where the model, the runtime, the silicon, the governance, and the productivity suite are all built by the same company and optimised to work together. It is a large bet. Cloud Next 2026 is where Google is asking enterprises to take it.