The concept began in November 2025 with a project internally codenamed Pi. The core philosophy was radically simple: an agent shouldn't be a pre-programmed tool, but a "malleable harness" for an LLM.
The "Tiny Core" Philosophy:
- Minimalism: The harness was designed to be as small as possible, providing only the bare essentials: a shell (
exec), file access (read/write), and a way to persist state. - Self-Modification: Pi was one of the first implementations to allow an agent to write its own plugins and modify its own "Soul" (system prompt) during a session.
- RL into Reality: By saving its successes and failures into local memory files, the agent would effectively "Reinforcement Learn" its way into the specific persona and toolset the user required.