Google's release of Managed Agents in the Gemini API is the signal to pay attention to this week. It packages the messy, stateful, and insecure parts of building agents into a single API endpoint, backed by a new, cost-effective frontier model, Gemini 3.5 Flash. The takeaway is that the infrastructure for running autonomous agents in secure, isolated environments is now a utility.
what actually shipped
On May 19, 2026, Google released two things that matter for builders: Gemini 3.5 Flash and the public preview of Managed Agents for the Gemini API. Gemini 3.5 Flash is positioned as a model optimized for performance on agentic and coding tasks. It's the engine.
The more significant release is Managed Agents. This is the platform. It gives developers the ability to build and deploy autonomous, stateful agents that run in secure, Google-hosted Linux sandbox environments. Instead of managing your own infrastructure for code execution and state, you can now spin up an agent via an API call. The first available general-purpose agent is antigravity-preview-05-2026, which can plan, reason, write and execute code, manage files, and browse the web inside its container.
from a local cli to a server-side platform
This release coincides with a strategic shift. Google is transitioning its popular Gemini CLI to a new Antigravity CLI. This isn't just a rename. It reflects a move from a local terminal utility to a client for a unified, server-side agent platform. The new CLI is built in Go for better performance and supports asynchronous workflows, letting you orchestrate multiple agents on complex tasks without locking your terminal.
This transition acknowledges that real agentic work involves multiple agents and shared context, which outgrew the initial CLI's scope. By unifying the backend into the Antigravity platform, improvements to the core agent harness are automatically available to the CLI, the desktop app, and the API. For developers, this means the agent you prototype in the terminal shares the same foundation as the one you deploy to production.
how you will use this
Building with this new API means focusing less on the infrastructure of agent execution. You are no longer primarily responsible for the security of running model-generated code or persisting state between long-running tasks. You define the task and the tools, and the managed agent handles the execution loop within its sandboxed environment.
A request to the new Interactions API might look conceptually like this. You provide the model, the agent definition, and the user's high-level task, and the platform manages the multi-step execution.
from google.ai import agents
# Configure the managed agent with a specific toolset and model
file_processing_agent = agents.ManagedAgent(
name="projects/my-project/agents/file-processor",
model="gemini-3.5-flash",
# The platform provides the secure code execution environment
harness="antigravity-preview-05-2026",
tools=[
agents.Tool(name="file_reader"),
agents.Tool(name="data_transformer"),
agents.Tool(name="report_generator"),
]
)
# Start a stateful session to perform a multi-step task
session = file_processing_agent.start_session()
# The agent plans and executes steps inside its isolated sandbox
final_result = session.execute("Analyze the quarterly sales data in /uploads, identify the top three regions, and generate a PDF summary.")
print(final_result.get_file("sales_summary.pdf"))
The key change is moving from a request-response loop that you manage to a persistent, stateful agent that you task. The antigravity-preview-05-2026 agent harness provides the core capabilities of file management, web browsing, and code execution out of the box.
the so-what for builders
The move toward managed, server-side agents is a significant abstraction layer. For the last year, building a truly autonomous agent meant wrestling with Docker containers, file system permissions, and state management. Google is now offering to handle that plumbing. This lowers the barrier to entry for shipping sophisticated agentic workflows.
This doesn't eliminate the hard problems of agentic reasoning and reliability. But it does commoditize the execution environment, letting you focus on the agent's actual logic and purpose. It's a platform bet that the future of AI development is less about prompting a model and more about directing a stateful worker.
Top comments (0)