Create an Open Deep Research Multi-Agent in Python (Step by Step)

Build a complete multi-agent deep research system with open-source models: use Hugging Face Inference Providers, Firecrawl MCP tools for web search and scraping, and smolagents for agent coordination. --- 🔗 Links - Create Firecrawl account: https://firecrawl.link/alejandro-ao - Create a Hugging Face account: https://huggingface.short.gy/join --- 🤗 Connect with me - X: https://x.com/_alejandroao - LinkedIn: https://www.linkedin.com/in/alejandro-ao/ --- 🔗 Code - Colab Notebook: https://colab.research.google.com/gist/alejandro-ao/124673ecfc2e860cbb592e9cfeeae256/deep-research.ipynb - GitHub Repo: https://github.com/alejandro-ao/open-deep-research-w-firecrawl - Written guide: https://alejandro-ao.com/posts/agents/multi-agent-deep-research/ 🔗 *Resources* - Anthropic's article: https://www.anthropic.com/engineering/multi-agent-research-system - Intro to Agents: https://youtu.be/vHDwpoSFdQY --- ⏰ *Timestamps* 00:00:00 Introduction to Deep Research System Tutorial 00:02:38 Setting Up Smolagents and Integrations 00:05:03 Generating the Research Plan with Models 00:09:24 Generating Subtasks with Research Coordinator 00:17:35 Understanding the Subtask JSON Structure 00:19:25 Choosing Models for Coordinator & Sub-Agents (Context Window & Tools) 00:22:03 Defining the run_deep_research Orchestrator 00:23:23 ToolCallingAgent Basics & Firecrawl MCP Overview 00:26:06 Loading Firecrawl MCP Tools into Smolagents 00:28:05 Implementing the initialize_subagent Tool & Sub-Agent Prompt 00:30:42 Building the Coordinator Agent & Coordinator Prompt 00:34:27 Running the Full Deep Research Pipeline (Coordinator + Sub-Agents) 00:37:08 Reviewing Results of Deep Research Execution 00:41:32 Closing Remarks and Future Topics