Topic

DeerFlow 2.0: From Deep Research Framework to Super Agent Runtime – The Path of Architectural Refactoring

As AI agents evolve from “instant Q&A” to “deep thinking,” a fundamental challenge emerges: how can we move beyond second-level responses and enable AI to complete complex tasks that span minutes or even hours? DeerFlow 2.0 offers a practical answer. In this talk, we will dive into how we re-architected DeerFlow from a 1.x deep research framework into a Super Agent runtime, featuring sub-agent collaboration, hierarchical memory, and sandboxed execution. We will also explore how to strike a balance between security and performance, along with key insights into real-world agent implementation. Whether you are a technical decision-maker, developer, or researcher interested in the future of AI agents, this session will provide both hands-on experience and deep perspectives on building next-generation AI applications. Outline Why a Rewrite? Limitations of DeerFlow 1.x and the architectural vision behind 2.0 Core Technical Decisions LangGraph, MCP, sandboxing, and hierarchical memory Engineering in Practice Security auditing, performance optimization, and sub-agent collaboration Rethinking Agents Tools, memory, collaboration, intelligence, and emotional intelligence Community & Roadmap Open-source practices behind 55K stars Key Takeaways 1. Understanding the Value of DeerFlow More than a deep research tool — a general-purpose agent runtime Layered architecture, sub-agent collaboration, and long-term memory 2. Mastering Core Principles of Agent Development Tool standardization (Skills / Tools / MCP) Persistent memory through layered design Security-first approach with sandbox isolation 3. How to Get Involved Contribute code, documentation, and Skills Become an active member of the open-source community