The contest among the big AI labs has quietly changed shape. The question is no longer whose model writes the cleanest paragraph but whose platform can run software agents inside a company's own systems. That shift, from answering questions to taking actions, is what makes this round of announcements matter for the people whose work those systems touch.
Google bets on distribution
At Google Cloud Next 2026 the company rebranded Vertex AI as the Gemini Enterprise Agent Platform and launched Workspace Studio, a no code builder for assembling AI workflows. The platform now exposes more than 200 models, including third party systems such as Anthropic's Claude, alongside a web browsing agent called Project Mariner, managed MCP servers across Google Cloud, and a production version of the Agent2Agent protocol that lets agents from rival vendors talk to one another. The bet is reach over raw benchmark scores: an adequate agent wired into tools a billion people already use can outweigh a stronger one sold on its own.
OpenAI goes after the back office
OpenAI is pushing in the same direction from a different entry point. Its Operator agent is reported to score 87 percent on complex browser task benchmarks, and the company has recruited integrators including Cognizant and CGI to carry its Codex coding agent into enterprise software shops. Enterprise revenue now reportedly accounts for around 40 percent of OpenAI's total, a sign that the money is moving from consumer subscriptions to corporate deployments.
Anthropic adds the plumbing, and a warning
Anthropic has focused on the infrastructure that makes agents safe to run on private data, adding MCP tunnels for private network access and self hosted code execution sandboxes to its Claude Managed Agents platform. At the same time, on 4 June the company published a piece titled When AI builds itself, arguing for a globally coordinated slowdown on frontier development, an unusual note of caution from a firm shipping the very tools it warns about.
What it means for work
When an agent can act across a company's applications rather than answer a single prompt, the unit of automation stops being a task and becomes a workflow. The roles most exposed are the ones largely defined by moving information between systems. The durable human contribution shifts toward designing, supervising and auditing what the agents do, which is why fluency with these platforms is fast becoming a baseline expectation rather than a specialist edge.
