What are Managed Agents in the Gemini API and how do they differ from building agents yourself?

from google import genai
client = genai.Client()

# Invoke a managed agent exactly like a model call:
interaction = client.interactions.create(
    model="antigravity-preview-05-2026",  # managed agent
    input="Build and test a Python REST API with FastAPI.",
)

# The agent handles:
# 1. Planning the approach
# 2. Writing the FastAPI code
# 3. Installing dependencies (pip install fastapi uvicorn)
# 4. Running tests
# 5. Reporting results
# All inside its sandbox - zero infrastructure from you

# You can also mix agents and models in one conversation:
research = client.interactions.create(
    model="gemini-deep-research-preview",  # Deep Research agent
    input="Research best practices for API authentication",
)

# Follow up with a standard model using the research context:
followup = client.interactions.create(
    model="gemini-3.5-flash",
    input="Now implement authentication based on this research.",
    previous_interaction_id=research.id,  # carries context forward
)