Antigravity (Gemini Agents API)

Examples for integrating Google's Gemini Agents API ("Antigravity") with Agno.

Antigravity is a managed agent loop: a single REST call spins up a sandboxed Linux environment with web search, code execution, and file I/O built in, runs an autonomous loop to satisfy the request, and returns or streams events. Environments persist across turns via an environment_id.

Agno integrates Antigravity two ways. Pick based on who drives the loop:

Integration	When to use	Example
`AntigravityAgent` (external agent)	You want Antigravity to be the agent — served through AgentOS, with Agno handling sessions, streaming, and the UI.	`antigravity_basic.py`, `antigravity_session_agentos.py` (this folder)
`AntigravityTools` (toolkit)	You want a regular Agno agent (any model) to delegate a sub-task to an Antigravity sandbox as one tool call.	`cookbook/91_tools/antigravity/antigravity_tools.py`

Setup

bash

export GEMINI_API_KEY=...

Files

antigravity_basic.py — minimal standalone run
antigravity_session.py — multi-turn session, environment reused across turns
antigravity_sources.py — pre-load files into the sandbox (inline / GCS / repo)
antigravity_custom_agent.py — register a named custom agent (Agents API) and invoke it
antigravity_from_agent_directory.py — load an agent from a local agent.yaml + AGENTS.md + workspace/ + skills/ folder (see example_agent/)
antigravity_snapshot.py — download an environment's filesystem as a tar archive
antigravity_session_agentos.py — same with SQLite-backed sessions

For the toolkit examples (Antigravity-as-a-tool), see:

cookbook/91_tools/antigravity/antigravity_tools.py — delegate a sub-task to a sandbox
cookbook/91_tools/antigravity/antigravity_agents_crud_tools.py — full Agents API CRUD (create / get / update / list / versions / delete / invoke)
cookbook/91_tools/antigravity/antigravity_directory_tools.py — wire a local agent.yaml folder into the toolkit (parses, registers, routes future tool calls at the named agent)
cookbook/91_tools/antigravity/antigravity_snapshot_tools.py — have an Agno agent run a sandbox task then download the resulting environment as a tar

Test results are tracked in TEST_LOG.md.