How Gemini Managed Agents Works under the Hood
How Gemini Managed Agents Works under the Hood
June 10, 20264 minute read
You can go from zero to a working agent in five lines. One API call, and the response comes back with a finished PDF, charts, and a summary.
Python
from google import genai
client = genai.Client()
interaction = client.interactions.create( agent="antigravity-preview-05-2026", input="Analyze the latest lego set data and create a PDF report with charts", environment="remote", )
print(interaction.output_text)
What most people miss: this is not a single model call that returns text. Behind that one call, a sandbox boots, skills get loaded, and the model enters a loop where it reasons, picks tools, executes code, reads the output, and repeats until the task is done.
The Execution Loop, Step by Step
interactions.create() does more than send a prompt. It spins up a full execution environment.
User
Your code
Gemini API
Orchestrator
Gemini 3.5 Flash
Reasoning
Sandbox
Linux VM
- You send a prompt to the Gemini API
- The API boots a sandbox, routes your prompt to the model, dispatches tool calls, and collects results
- Gemini 3.5 Flash plans and picks tools, looping until the task is complete
- Code runs inside an isolated Linux VM, not on your machine
Simulate the flow
requestresponseexecuteresults
User
Write a python script that prints hello world and execute it.
Gemini API
code_execution
filesystem
Skills
0 loaded
Gemini 3.5 Flash
Sandbox
offline
rootfs
Terminal
Ready
Debian 4 vCPU 16G VM
GCS
Repository
Inline
Prompt
Write a python script that prints hello world and execute it.
Simulate
Script created and executed. - File: hello.py - Output: hello world
Explore each step
1Request2Boot Sandbox3Mount Sources4Discover Skills5Reason6Tool Call7Execute8Return Results9Loop10Final Response
Request : Your prompt hits the Interactions API via HTTPS. The API accepts it and starts orchestrating the interaction server-side.
requestresponseexecuteresults
User
Analyze sales data and create a PDF report
Gemini API
code_execution
filesystem
Skills
0 loaded
Gemini 3.5 Flash
Sandbox
offline
rootfs
/workspace
Terminal
Isolated Linux VM · 4 vCPU · 16 GB
Debian 4 vCPU 16G VM
GCS
Repository
Inline
Previous1 / 10Next
Inside the Sandbox
Each interaction gets its own isolated Linux container. The agent can install packages, run scripts, read and write files, and browse the web. The sandbox runs Ubuntu with 4 vCPU and 16 GB RAM. Environment compute is not billed during preview (you pay only for model tokens).
Environments can persist across interactions. The first call returns an environment_id you pass back to keep files and state intact. See the Environments documentation for sources, networking, and lifecycle details.
Python
Turn 1: sandbox is created
i1 = client.interactions.create( agent="antigravity-preview-05-2026", input="Install pandas and create analysis.py", environment="remote", )
Turn 2: same sandbox, files persist
i2 = client.interactions.create( agent="antigravity-preview-05-2026", input="Run analysis.py on the Q1 data", environment=i1.environment_id, previous_interaction_id=i1.id, )
From Prompt to Production Agent
The execution loop is the same whether you are experimenting or running in production. The difference is how much configuration you want to provide upfront.
- Ad-hoc call: Send a input with
environment="remote".
Python
interaction = client.interactions.create( agent="antigravity-preview-05-2026", input="Write a Python script that prints hello world", environment="remote", )
- Add instructions and skills: Use
system_instructionorAGENTS.mdto customize behavior and provide skills viasources.
Python
interaction = client.interactions.create( agent="antigravity-preview-05-2026", input="Analyze Q1 revenue and create a slide deck.", system_instruction="You are a data analyst. Export results as PDF.", environment={ "type": "remote", "sources": [ {"type": "inline", "target": ".agents/AGENTS.md", "content": "Always use matplotlib for charts."}, {"type": "inline", "target": ".agents/skills/slides/SKILL.md", "content": "---\nname: slides\n---\n# Slide Maker\nCreate HTML decks."}, ], }, )
- Save as a managed agent: Once your setup is stable, persist it with
client.agents.create(). Each invocation forks a fresh sandbox from the base environment.
Python
agent = client.agents.create( id="data-analyst", base_agent="antigravity-preview-05-2026", system_instruction="You are a data analyst. Export results as PDF.", base_environment={ "type": "remote", "sources": [ {"type": "repository", "source": "https://github.com/my-org/templates", "target": "/workspace/templates"}, ], }, )
- Invoke by ID: Reference the agent ID on every call. Same loop, no config to repeat.
Python
result = client.interactions.create( agent="data-analyst", input="Analyze Q1 revenue data and create a slide deck.", environment="remote", ) print(result.output_text)