Skip to content

autoagent-tutorial

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

autoagent-tutorial

README.md

layout default
title AutoAgent Tutorial
nav_order 140
has_children true
format_version v2
source_repo https://github.com/HKUDS/AutoAgent
categories
ai-agents
zero-code
multi-agent
deep-research
related_tutorials
autoresearch-tutorial
openhands-tutorial
agno-tutorial
crewai-tutorial
last_updated 2026-04-12

AutoAgent Tutorial

AutoAgent (formerly MetaChain) is a zero-code autonomous agent framework from HKUDS that lets you describe agents in plain English and have them generated, tested, and deployed automatically. With 9,116 GitHub stars and an academic paper (arxiv:2502.05957), it represents a significant step toward democratizing multi-agent system development.

Why This Track Matters

AutoAgent is useful because it turns agent development into an interactive generation workflow. Instead of requiring a developer to hand-wire every tool, prompt, and orchestration step, it exposes modes for research, agent creation, and workflow creation that beginners can reason about separately.

What You Will Learn

This tutorial walks through AutoAgent from first install to production-grade multi-agent pipelines. By the end, you will understand how the MetaChain engine works under the hood, how all three operating modes fit together, and how to extend the framework with your own tools, agents, and workflows.

Current Snapshot (auto-updated)

Who This Tutorial Is For

  • Developers who want to build research or automation agents without writing orchestration boilerplate
  • ML engineers evaluating AutoAgent for benchmarks (GAIA, Math500, Agentic-RAG)
  • Contributors looking to add tools, agents, or new evaluation suites to the ecosystem

Naming Note

The internal codebase uses the class name MetaChain throughout — the project was publicly renamed from MetaChain to AutoAgent in February 2025. You will see from autoagent import MetaChain and MetaChain.run() in all source files. This tutorial uses "AutoAgent" when referring to the product and "MetaChain" when referring to the specific class or import.

Mental Model

Think of AutoAgent as three coordinated workbenches. User Mode runs research tasks, Agent Editor turns natural-language requirements into reusable agents, and Workflow Editor connects agents into event-driven pipelines.

Three Operating Modes

Mode Entry Point Best For
User Mode (Deep Research) auto main Open-ended research, file analysis, web browsing
Agent Editor auto main → "create agent" Generating new agents from NL descriptions
Workflow Editor auto main → "create workflow" Composing async parallel pipelines

Chapter Guide

  1. Getting Started — Install, .env setup, first research task, three-mode overview
  2. Core Architecture: MetaChain Engine — Agent/Response/Result types, run loop, context_variables, non-FC XML fallback
  3. The Environment Triad — DockerEnv TCP server, BrowserEnv Playwright, RequestsMarkdownBrowser
  4. User Mode: Deep Research System — SystemTriageAgent, agent handoff, multimodal web surfing, GAIA benchmark
  5. Agent Editor: From NL to Deployed Agents — 4-phase pipeline, XML form schema, ToolEditorAgent, AgentCreatorAgent
  6. Workflow Editor: Async Event-Driven Pipelines — EventEngine, listen_group(), GOTO/ABORT, parallel execution
  7. Memory, Tool Retrieval, and Third-Party APIs — ChromaDB ToolMemory, LLM reranker, RapidAPI ingestion, token budget
  8. Evaluation, Benchmarks, and Contributing — GAIA, Math500, Agentic-RAG, adding benchmarks, contributing tools/agents

Architecture at a Glance

flowchart TD
    U[User] --> CLI["auto main CLI"]
    CLI --> UM[User Mode / Deep Research]
    CLI --> AE[Agent Editor]
    CLI --> WE[Workflow Editor]
    UM --> MC["MetaChain Engine (core.py)"]
    AE --> MC
    WE --> EE["EventEngine (flow/)"]
    MC --> DE["DockerEnv\n(TCP :12346)"]
    MC --> BE["BrowserEnv\n(Playwright)"]
    MC --> MB["RequestsMarkdown\nBrowser"]
    MC --> REG["Registry\n(tools/agents/workflows)"]
Loading

Quick Start

git clone https://github.com/HKUDS/AutoAgent
cd AutoAgent
pip install -e .

# Set up .env with your provider keys
cp .env.example .env
# Edit .env: OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.

auto main

Key Technical Facts

Property Value
Language Python 3.10+
License MIT
LLM routing LiteLLM 1.55.0 (100+ providers)
Code isolation Docker (tjbtech1/metachain image, TCP port 12346)
Memory/retrieval ChromaDB + sentence-transformers
Browser automation Playwright + BrowserGym
Stars 9,116
Paper arxiv:2502.05957

Source References

Navigation