Google’s Python SDK for Gemini agents ships with a compiled catch

What it does

The Google Antigravity SDK is a Python library for building AI agents on top of Gemini. It wraps the full agent lifecycle—binary discovery, tool registration, hook execution, and stateful session management—inside a high-level Agent class and a lower-level Conversation API. You feed it prompts, custom Python functions, multimodal files, or external MCP servers, and it handles the back-and-forth with the model. The whole thing is async and streams tokens, reasoning steps, and tool calls in real time.

The interesting bit

The SDK distributes a compiled runtime binary inside its PyPI wheels, so cloning the repository is intentionally insufficient to run it—a curious packaging choice for an open-source Python library. It also defaults to read-only mode and exposes a declarative policy engine, which is more safety scaffolding than you usually see in agent frameworks.

Key highlights

• Agent and Conversation APIs cover both quickstarts and low-level session control.
• Native streaming of tokens, reasoning steps, and ToolCall events.
• Multimodal support for images, video, audio, and documents from memory or disk.
• Built-in MCP integration and a declarative policy layer that defaults to read-only.
• Background triggers for pushing external events into the agent loop.

Caveats

• Requires installing from PyPI; the compiled runtime binary is not present in the source repository, so you cannot run the SDK from a raw clone.

Verdict

This is a batteries-included scaffold for developers building on Gemini who want safety policies, streaming, and multimodal ingestion. If you need a fully source-available runtime or are not in the Gemini ecosystem, it is a poor fit.