The defining contest in AI this year is no longer which model scores highest on a benchmark, but whose agents can reliably do a day's office work. At Cloud Next 2026 Google made its most aggressive move yet, rebranding Vertex AI as the Gemini Enterprise Agent Platform and folding its scattered tools into a single product aimed squarely at the rivals it has been chasing.
Google's pitch: own the whole stack
The centrepiece is Workspace Studio, a no code builder that lets ordinary office staff describe an agent in plain language and deploy it across Gmail, Docs, Sheets, Meet and Chat, with hooks into third party apps such as Salesforce, Jira and Asana. The platform's Model Garden now offers more than 200 models, including Anthropic's Claude, and DeepMind's web browsing agent Project Mariner scores about 83.5 percent on the WebVoyager benchmark while handling ten tasks at once. Chief executive Thomas Kurian framed the strategy bluntly, arguing that competitors are handing customers the pieces, not the platform, while Google controls everything from its Ironwood chips to three billion Workspace users.
A protocol land grab
Underneath the products, the industry is quietly agreeing on plumbing. Google's Agent2Agent protocol, which lets agents from different vendors talk to one another, has reached around 150 organisations in production with backing from Microsoft, AWS, Salesforce, SAP and ServiceNow, and it sits alongside Anthropic's Model Context Protocol rather than against it. Standardising how agents communicate matters because it turns isolated assistants into a workforce that can hand tasks between systems, which is precisely the capability that makes automation of routine office work plausible.
OpenAI and Anthropic press their own fronts
The challengers are not standing still. OpenAI's Operator agent is reported to score about 87 percent on complex browser task benchmarks, and the company has enlisted consultancies including Cognizant and CGI to push its Codex coding agent into enterprise software teams, with enterprise work now said to account for roughly 40 percent of its revenue. Microsoft and Google have both moved onto Anthropic and OpenAI's turf in coding models, turning what was a research rivalry into a direct fight over the most automatable, highest paid corner of the office.
For workers the through line is unmistakable. Every one of these launches is designed to let software take actions across the applications where knowledge work actually happens, not merely answer questions about it. The labs frame this as productivity, and for many roles it will be. But an agent that can read an inbox, update a CRM and draft the follow up is, by another name, a description of several junior jobs. The infrastructure being assembled this year is the precondition for the workforce shifts the rest of the industry keeps forecasting.
