GitHub - openai/openai-agents-js: A lightweight, powerful framework for multi-agent workflows and voice agents
A lightweight, powerful framework for multi-agent workflows and voice agents - openai/openai-agents-js
Personal Notes
I got to meet Guido van Rossum and Steve Lucco (Python and Typescript founders, respectively) at the PyBay 2025 conference yesterday and we chatted about TypeAgent and structured RAG.
AI Agent Research Project
Summary
TypeAgent is a Microsoft research project exploring an architecture for a personal agent that combines LLMs with traditional software. It introduces Structured RAG for precise conversational memory and the AMP system for natural information flow between actions and memories.
Highlights
- Introduces Structured RAG, a new indexing and query approach that substantially outperforms Classic RAG for recalling past conversation details
- Proposes three core principles: distill models into logical structures, use structure to control information density, and use structure to enable collaboration
- The AMP architecture enables actions and memories to flow together naturally, so past actions become parameters for future ones
- Authored by Guido van Rossum (Python creator) and Steve Lucco (TypeScript creator), giving it notable pedigree in the programming language community
- Aims to map user requests to actions at much lower cost and latency than current systems by distilling LLMs into logical systems
auto-generated
Context
Audience
AI researchers and developers working on LLM-based agent systems, particularly those interested in agent memory architectures and structured prompting approaches
DomainAI Agents and Language Models
FormatOpen source code repository with documentation
AccessFree, open source on GitHub
Source Overview
TypeAgent is sample code that demonstrates an architecture for a personal agent using language models and structured prompting to interact with various application agents.
It introduces Structured RAG for agent memory, enabling more precise recall of conversational history, and integrates action, memory, and planning into a unified flow via the AMP system. The repository includes example agents and a shell application for experimenting with natural language interfaces, but is intended for research and development purposes only, not for production use.
Topics
Related
Structured RAGAMP (Action Memory Planning)Tree of ThoughtClassic RAGPersonal AgentsStructured Prompting
