Gemini Managed Agents: Deploy an AI Agent with One API Call

Google shipped Managed Agents in the Gemini API at I/O 2026, and the pitch is blunt: call an API, get a working Linux sandbox where an AI agent reasons, runs code, browses the web, and manages files — all in one call. No VMs to spin up, no orchestration loop to write, no tool registry to maintain. The first call to client.interactions.create() replaces what used to be a multi-service setup. Whether or not it holds up in production, the developer experience argument is hard to argue with.

What Managed Agents Actually Do

The core of this is the Interactions API (currently in Beta), which replaces generateContent for agentic work. You pass a task to the Antigravity agent — antigravity-preview-05-2026, Google’s general-purpose managed agent — and Google provisions a remote Linux sandbox, runs the agent loop, and returns the result. Inside that sandbox, the agent can execute Python, Node.js, and Bash; install packages; read and write files; and browse the web. That last combination — code execution and web browsing in the same isolated environment — is what separates this from OpenAI’s code interpreter, which keeps those capabilities more siloed.

The Antigravity agent runs on Gemini 3.5 Flash, fine-tuned specifically for tool calls and shell command output. Google reports Terminal-Bench 2.1 at 76.2% and MCP Atlas at 83.6%, outperforming Gemini 3.1 Pro on the agentic benchmarks that matter for this use case.

The Code: Start in Under Five Minutes

Install the SDK (pip install google-genai), set your GEMINI_API_KEY, and you’re ready. Here’s the basic pattern:

from google import genai

client = genai.Client()

# One call provisions a fresh Linux sandbox
interaction = client.interactions.create(
    agent="antigravity-preview-05-2026",
    environment="remote",
    input="Research the top Python web frameworks in 2026. Write a markdown comparison and save it as report.md"
)

print(interaction.output_text)  # Final response
print(interaction.steps)        # Every reasoning step and tool call

# Continue in the same sandbox — files and packages persist
follow_up = client.interactions.create(
    agent="antigravity-preview-05-2026",
    environment=interaction.environment_id,
    previous_interaction_id=interaction.id,
    input="Add benchmark data and convert report.md to HTML"
)

The interaction.steps list is worth examining when debugging. It surfaces every reasoning step, tool invocation, and code execution so you can trace exactly what the agent did — significantly more useful than a black-box response.

State: Two Independent Dimensions

The state model is what makes iterative workflows practical. The Interactions API tracks two independent dimensions:

• Conversation context — tracked via previous_interaction_id. Pass the last interaction.id and the agent carries forward its full chat history, reasoning trace, and tool logs.
• Environment state — tracked via environment_id. Reuse the same sandbox and your installed packages, generated files, and file system state persist. Sandboxes have a 7-day TTL.

These two are independent by design. You can reset conversation history while keeping the environment — useful for running different prompts in the same prepared sandbox — or vice versa. Reusing environment_id is the better practice for iterative tasks: it avoids re-provisioning latency and keeps the packages you installed in step one alive for step three.

What It Costs Right Now

During the public preview, sandbox compute is free. You pay only for tokens at standard Gemini 3.5 Flash rates: $1.50 per million input tokens and $9.00 per million output tokens. That’s 25% cheaper than Gemini 3.1 Pro ($2.00/$12.00), which is now outclassed on agentic benchmarks anyway. When Managed Agents move to GA on the Enterprise Agent Platform, Google Cloud pricing applies. If you’re evaluating this for a real use case, the preview window is the right time to benchmark it.

Where This Fits (Honest Take)

Managed Agents are not the right tool for every agent workflow. Here’s where they earn their place and where they don’t.

Pick Managed Agents when: you want a working prototype in an afternoon, you’re already on GCP or deep in Google Workspace, or your use case is self-contained tasks — research, code execution, data processing — that don’t require tight integration with external services you control.

Stick with Claude Agent SDK or OpenAI when: you need fine-grained control over the orchestration loop, complex multi-agent topologies with MCP integration, or production-grade reliability guarantees that a preview product can’t provide. The Claude SDK gives you composable agents with explicit tool registration and transparent state management — more setup, more control.

The honest comparison: Managed Agents are to agent development what Vercel was to deployment — they remove the infra work so you can focus on what your agent actually does. That’s genuinely valuable. Not everyone needs more control than that.

Get Started

The official quickstart covers your first interaction in Python, JavaScript, and REST. The Antigravity agent reference has the full capability list. If you want to build a custom managed agent with your own tool set and system prompt, the documentation is at Building Managed Agents.

The preview is live now. Sandbox compute is free. If you’ve been meaning to prototype an agent workflow, this is the lowest-friction entry point available — with the understanding that preview means preview.