Minimal but runnable: the core loop is visible at a glance.
v1 sync loop: smallest while-loop agent for learning the basics.
v2 async: task decoupling + step state machine to avoid blocking requests.
Safety guardrails: dangerous-command blocks, timeouts, output truncation.
I built this because I got tired of configuring "Agent Frameworks" before I could actually write any logic. I wanted a workflow where the IDE itself handles the cognitive architecture.
Antigravity is a Python scaffold designed specifically for AI IDEs like Cursor or Google Antigravity.
The concept is "Zero-Config":
1. Git clone the repo.
2. Open it in Cursor.
3. The included `.cursorrules` file automatically "brainwashes" the IDE with a specific cognitive protocol (Infinite Memory + Tool Dispatch).
You don't need to setup the loop. You just tell the IDE: "Build a stock analysis agent," and because the scaffold is already loaded into the context, the IDE knows exactly how to fill in the blanks using the pre-built memory and tool interfaces.
It uses pure Python (no heavy dependencies) and is optimized for Gemini 2.0 Flash.
Happy to hear your thoughts on this "IDE-Native" approach!