DeerFlow 2.0 Architecture and Agent Design Deep Dive

DeerFlow 2.0 is introduced as a multi‑language project: the frontend is built with TypeScript on the React framework, while the backend is a Python service built on the LangGraph workflow framework and exposed via FastAPI REST endpoints.

The core capabilities reside in the deerflow‑harness package, which implements Agents, SubAgents, and Skills. Skills are loaded through the load_skills function; the current skill support is described as elementary and comparable to the Nanobot project.

Agent collaboration is offered in three modes: (1) LangGraph workflow style, (2) hierarchical parent‑child SubAgent structure, and (3) ACP‑protocol based agents. The system leverages LangGraph’s directed acyclic graph (DAG) and loop graph features to manage agent state transitions, and it supports spawning Sub‑Agents with specific system prompts and toolsets.

AgentMiddleware provides a plugin loading mechanism for various hooks. The LeadAgent initiates the workflow by loading context, spawning Sub‑Agents, and starting a loop cycle, as illustrated by the accompanying diagrams.

SubAgents are dynamically delegated rather than statically defined; their scheduling relies on an asynchronous thread pool combined with rate‑limiting and polling. SubAgents focus on task execution and are considerably simpler than the LeadAgent.

In summary, DeerFlow 2.0 closely mirrors OpenClaw’s Python capabilities but still lags in areas such as agent collaboration patterns, skill richness, and multi‑channel queuing. Nonetheless, adopting LangGraph as the workflow backbone proves effective for rapid development.