TypeAgent

TypeAgent is a Microsoft project providing sample code for a personal agent architecture. Its core goal is distilling LLM outputs into typed logical structures to build a single entry point that orchestrates multiple application agents.

The project is built on three design principles: distill models into logical structures (Actions discover translation patterns, Memory builds ontologies, Plans use Tree of Thought); use structure to control information density (discrete action categories, compact semantic structures, focused sub-problem nodes); and use structure to enable human-model-program collaboration (human disambiguation, simple model extraction, multi-model collaborative search expansion).

Core components include: TypeAgent Dispatcher (structured prompting + LLM routing to the best-matching agent), KnowPro (Structured RAG-based conversation memory system, significantly outperforming traditional RAG on retrospective queries), TypeAgent Cache (caching action translation patterns to reduce LLM call costs), and the AMP unified information flow architecture. Agent actions are defined via TypeChat Schema which also validates LLM responses, and action results automatically enter the memory system for future reference.

The project is organized as a monorepo with TypeScript (81.3%) as the primary language, plus Python, .NET, and Android subdirectories. Interaction modes include an Electron desktop Shell (with voice support), CLI, and a browser extension (allowing websites to register actions). It includes 16+ example agents covering music, calendar, email, VS Code, desktop control, image, markdown, montage, and more. The LLM backend is currently only validated with Azure OpenAI. State is stored locally as text/JSON files, telemetry is off by default, and only English has been tested. The project is in early active development under the MIT license.