At Google I/O, the company introduced Managed Agents in its Gemini API, a service that promises to combine weeks of agent deployment work into a single API call. It’s also a sign that Google believes its ecosystem, including the recently released Antigravity CLI, is ready to own the end-to-end execution layer.
Before a single agent is written, teams already spend days on unglamorous work: preparing runtime environments, managing sandboxes, wiring up tool call infrastructure. Model providers like Anthropic have launched platforms to handle much of that work, but Google’s approach is different.
Google said in a blog post that managed agents in the Gemini API “abstract complexity so you can focus on the product experience and agent behavior.” The service is available in preview through new custom templates in Google AI Studio.
The growth has introduced a real architectural question: should agent management live in the execution layer (built into the model or its harness) or in the infrastructure layer, as a separate runtime?
Comparing Google’s approach
Until recently, agent orchestration relied on frameworks that sat on top of the model, directing agents and allowing teams to control routing and execution separately. That layer is now being absorbed by the platforms themselves.
Recent platforms like Claude Managed Agents incorporate orchestration in the model layer instead of on a separate execution platform. The idea is that the model owns the reasoning and orchestration layers, and that companies have control over the execution.
AWS, through new capabilities in Bedrock AgentCore, add managed harnesses that tie together the initial tasks for agent deployment. Google’s approach goes further: it optimizes the model, harness, and sandbox together and runs everything in secure environments managed by Google.
Ramp’s René Sultan, quoted in Google’s announcement, said the change is concrete: "The real change with Gemini Managed Agents is that the agent runtime is moved to the platform. With the test environment, infrastructure, and execution cycle managed for you, developers can focus on producing the agent’s domain-specific behavior and iterate at a completely different pace."
The new reality of orchestration
Companies starting over with agents might find the platform offerings from Anthropic and Google strong, especially because they remove much of the difficulty of deploying agents while maintaining some control. Google, however, is committed to a more vertically integrated system, while Anthropic is committed to the model layer as an orchestration plane and AWS focuses on authorization.
But this also comes with some risks, according to XYO founder and CEO Arie Trouw.
“An additional risk is that developers will change what were previously deterministic services to what will now be probabilistic services, which can introduce unpredictable results for users at best, or data corruption at worst,” Trouw told VentureBeat in an email. “This is the classic example of having an awesome hammer and everything starts looking like nails. I’ve seen this pattern repeatedly as a business developer and founder over the past few decades.”
