PI Architecture EXPLAINED | Agent Loop, Tools, TUI and More

Pi operates through a minimalist, custom-built agentic loop that initializes context by integrating system prompts, agent files, and tool descriptions before executing language model calls. The system manages conversation history through a tree-structured JSONL format, allowing for efficient branching and bifurcation of sessions. Core functionality relies on four primary tools—read, bash, edit, and write—which can be extended via modular TypeScript packages that subscribe to specific workflow events. Compaction is handled dynamically by summarizing context based on token usage metrics, ensuring the model remains within its context window. The terminal user interface functions as a separate, component-based layer that intercepts commands and manages skill invocation, effectively separating the interactive experience from the underlying agent core logic.