Google has separated early agent development from manual cloud provisioning. At I/O 2026 on May 19, the company expanded its Antigravity platform and introduced Managed Agents in the Gemini API, two moves that together let a developer build, define and run a hosted agent on Google infrastructure starting from a Gemini API key. The same agent definitions then carry up to the Gemini Enterprise Agent Platform, a governed tier built for Google Cloud customers.

The split is the strategy. Google now runs two distinct lanes for agent builders. One serves individual developers, Google AI Studio users, Android builders and terminal-first developers who want to ship without provisioning cloud infrastructure. The other leads enterprise teams to governed deployment with identity controls, audit logs and policy enforcement. What connects them is a single agent harness, the runtime layer that handles reasoning, tool calls and code execution, which Google now ships across every surface.

That design choice gives Google a distinctive position against its two largest cloud rivals. Amazon and Microsoft both offer strong agent runtimes and developer tooling. What neither presents is the same kind of consumer-developer API on-ramp to a hosted agent runtime that is visibly continuous with its enterprise agent platform. For technology leaders deciding where their teams will build agents over the next two years, the question is whether a low-friction entry point is a genuine advantage or a funnel into a new form of platform dependence.

One Harness, Four Surfaces

Antigravity began as an agent-first development environment. Google has now widened it into what it calls an ecosystem with four ways to interact with the same underlying engine. Antigravity 2.0 is a standalone desktop application that acts as a home for orchestrating multiple agents in parallel, with scheduled tasks for background automation. The Antigravity CLI is a terminal-first surface for developers who want to spin up agents without a graphical interface. The Antigravity SDK gives programmatic access to the same harness and lets developers host agents on the infrastructure of their choosing. The fourth surface connects Antigravity directly to Google Cloud projects through the Gemini Enterprise Agent Platform.

Managed Agents extends that harness one step further. With a single Gemini API call, a developer can spin up an agent that reasons, executes code and browses the web inside an isolated, ephemeral Linux environment that Google operates. Each interaction creates a persistent environment that can be resumed in later calls with files and state intact. Developers define agent behavior declaratively, writing instructions and skills into markdown files named AGENTS.md and SKILL.md rather than building orchestration code by hand.

The detail that matters for the audience question is the on-ramp. A developer can start in Google AI Studio with a Gemini API key, without manually opening the Cloud console, configuring identity roles or enabling billing for the initial free-tier path. The key is still associated with a Google Cloud project that Google creates in the background, but the builder never touches it. Managed Agents is reachable from that same key. The friction is not eliminated so much as deferred.

How Amazon And Microsoft Compare

Amazon has the equivalent runtime layer. Bedrock AgentCore is a framework-agnostic, model-agnostic environment for deploying and operating agents, with managed memory, identity, gateway, code interpreter and observability. Amazon also ships Kiro , an agentic development environment available as an editor and a command-line tool. The difference is coupling. AgentCore is an AWS service for deploying agents at scale, and its hosted runtime is anchored in an AWS account with identity roles and storage created on the developer's behalf. The capability is strong, but the path to a hosted agent runs through the cloud account, and the tooling spans two product brands a developer assembles themselves.

Microsoft has the Foundry agent service , a fully managed platform for building, deploying and scaling agents that handles hosting, identity, observability and enterprise security. It is an Azure and Microsoft Foundry service, not a consumer-style API on-ramp aimed at developers who have made no cloud commitment.

So the contrast is real, though narrower than a winner-take-all reading suggests. All three vendors can host production agents, and all three have credible developer tooling. Google's differentiation is the continuity between a Gemini API and AI Studio entry point and the enterprise platform, running on the same agent definitions and harness, rather than the mere existence of agent tooling.

Where The Strategy Is Still Unproven

Several constraints temper the picture. Managed Agents is rolling out in preview, and Google’s documentation describes the Antigravity agent and the Interactions API as preview, with features and schemas subject to change. Enterprise support for Managed Agents on the Gemini Enterprise Agent Platform is in private preview, so the most governed end of the funnel is not yet generally available. A preview runtime is suited to experimentation and pre-production evaluation, not to production commitments.

The harness consolidation has also created friction. Google is encouraging Gemini CLI users to migrate to the Antigravity CLI, and Gemini CLI service for free, Pro and Ultra users is scheduled to stop on June 18, 2026, while enterprise access continues under paid Gemini and Gemini Enterprise Agent Platform keys. Developers who built workflows on the older command-line tool now face a forced move, and consolidation of that kind tends to surface complaints about usage limits and changed defaults before it settles.

The deeper open question is commercial. A low-friction on-ramp lowers the barrier to starting, but Google still needs those developers to convert into Google Cloud customers for the strategy to pay off. A funnel only works if builders move along it. Whether a developer on the free-tier path becomes an enterprise buyer is a behavioral bet, not a settled outcome.

What Technology Leaders Should Weigh

For enterprises evaluating where agent development happens, Google's split changes the early calculus. Teams can prototype and validate agent designs without manually provisioning infrastructure, then carry the same AGENTS.md and SKILL.md definitions into a governed environment when a project warrants it. That continuity reduces the rebuild cost that often appears when a proof of concept graduates to production.

The trade-off is concentration. Building on a single vendor's harness across every surface means the harness becomes the dependency, regardless of which entry point a team starts from. Leaders who valued multicloud flexibility in the infrastructure era should ask whether they are comfortable with the agent runtime becoming the new lock-in layer. The practical move is to treat the low-friction on-ramp as an evaluation channel rather than a default commitment, and to test agent portability before the harness becomes load-bearing.