Managed Agents on Gemini API show Google's ambition to sell infrastructure, not just model
Google's most telling move in AI isn't always in the model with the best demo. Sometimes it's in the layer that turns a model into a programmable product. This is what appeared with the Managed Agents announcement in Gemini API, published on May 19, 2026, but still very current in this week's adoption cycle. The message is clear: Google wants developers to buy not just intelligence, but the agent's operational infrastructure. ## What happened According to the official post, the Gemini API now supports managed agents. With a single call, the developer can launch the Antigravity agent in a secure and ephemeral Linux sandbox, with the ability to reason, use tools, browse the web, execute code, and maintain state between interactions. The system also allows you to define your own agents with instructions, skills and data, recorded in markdown files such as AGENTS.md and SKILL.md. In addition, the feature arrives in preview in both Gemini API and Google AI Studio, with templates to get started quickly, and gains enterprise support on the Gemini Enterprise Agent Platform. This shows two simultaneous fronts: convenience for individual developers and platform strategy for companies. It's not just a new SDK. It's an attempt to package the "agent runtime" as a service. ## The technique behind Building a production agent has always been more complicated than making a smart prompt. It is necessary to isolate execution, manage files, preserve sessions, control tools, deal with browsing and still maintain security. What Google is saying is: stop putting up that scaffolding manually; rent ours. Antigravity acts as a managed harness, provisioning a remote Linux environment, resumable state, and common action primitives. This has profound technical implications. By exposing AGENTS.md and SKILL.md as versionable artifacts, Google brings the definition of agents closer to the modern development flow: readable text, versioning, and iterability. At the same time, it abstracts an expensive part of the operation, especially for teams that don't want to invest early in their own sandbox infrastructure, observability and session continuity. There is a tension here. The more the provider simplifies the runtime, the more the developer gains speed. But dependence on that vendor's operational semantics also grows: session format, sandbox limits, navigation, tooling policy, cost and governance. That's the real game of the ad. The model matters, of course. But the potential lock-in is at runtime. ## Why this matters For startups and small teams, managed agents can compress months of work into days. Instead of building an isolated environment, file control and context resumption from scratch, the team starts with the product. This reduces barrier to entry and speeds up serious prototyping. For companies, the story is different: the value lies in standardization, control and scalability, especially when several agents need to live with corporate politics. There is also a competitive effect. Until recently, many providers sold models and let the ecosystem figure out the rest. Now, the big players are moving up the pile. Whoever controls the agent's runtime controls most of the cost, telemetry, skill standards and the path to production. Google is explicitly saying it wants to compete for this tier. ## The future it anticipates The plausible future is a market divided between "self-hosted" agents, with more control and higher operational costs, and "managed" agents, with more speed and more dependence on the supplier. In many cases, the hybrid route should win: prototype with managed runtime and insource critical parts as the product matures. It is also likely that AGENTS.md and SKILL.md will evolve from a practical convention to a unit of interoperability between tools. If this format, or something similar, crosses providers, the gain for developers will be enormous. If it stays stuck in a single-player ecosystem, it will remain useful but less transformative. ## What to watch out for The next few months will answer three central questions. The first is actual performance: how reliable is the sandbox on long streams, with browsing and persisted files? The second is economics: does the managed runtime reduce total cost or just shift infrastructure costs to platform costs? The third is portability: to what extent can an agent defined here migrate to another stack without heavy rewriting? The good reading of the ad is not "Google has launched another agent". It's "Google wants to be the invisible operating system for third-party agents." If you can combine speed, governance, and sufficient interoperability, this layer could be worth more than any single model benchmark.
Sources
- https://blog.google/innovation-and-ai/technology/developers-tools/managed-agents-gemini-api/
- https://ai.google.dev/