A teaching-oriented multi-agent framework (PicoAgents) with a companion book, covering the full path from building LLM agents from scratch to production deployment, with 50+ examples, DAG workflow engine, autonomous orchestration, Computer Use Agent, and evaluation framework.
This project is initiated by Victor Dibia, a core contributor to Microsoft AutoGen, and consists of two core components: the PicoAgents framework and the code repository for the companion book.
At the framework level, PicoAgents provides full Agent building capabilities including 15+ built-in tools with custom tool support, long-term memory and RAG, Pydantic-based structured output, native SSE streaming, extensible middleware, and OpenTelemetry observability. For multi-agent orchestration, it supports a type-safe DAG workflow engine (sequential/parallel with checkpoint recovery) and three autonomous orchestration modes—Round-Robin, AI-driven speaker selection, and Plan-based orchestration (Magentic One pattern)—with 9 built-in termination conditions. Additionally, it offers a Playwright-based Computer Use Agent (multimodal visual reasoning + browser automation) and a complete evaluation framework (LLM-as-Judge, reference matching, composite scoring). Production-grade optimizations include two-stage filtering for 90% cost reduction and a Think Tool for 54% performance improvement.
On the teaching side, the repository provides 50+ runnable examples organized by chapter, a four-step progressive minimal Agent implementation (code_along), course materials, and research references. The Web UI can be launched with a single command and auto-discovers Agents and Workflows in the current directory. The framework follows a framework-agnostic design principle—core patterns are transferable to LangGraph, AutoGen, Google ADK, etc., with equivalent implementations provided for comparison in the repository.
Supports OpenAI, Azure OpenAI, Anthropic, GitHub Models, and any OpenAI-compatible endpoint. Licensed under Apache-2.0.