DeerFlow 2.0 (SuperAgent harness)

Description. DeerFlow 2.0 is a ground-up rewrite of ByteDance's DeerFlow that shares no code with the 1.x deep-research graph and became the active line after its launch around late February 2026. Rather than a fixed LangGraph research workflow, it is a general open-source SuperAgent harness: a lead agent receives the request, decomposes it, and dynamically spawns subagents on the fly, each with its own scoped context, tools, and termination conditions. Subagents run in parallel when possible and report structured results back to the lead agent, which synthesises them into the final deliverable. The roster is not fixed: built-in subagents (a general-purpose agent and a bash agent) sit alongside user-defined custom agents declared in config.yaml, and concurrency is bounded by a SubagentLimitMiddleware. Each subagent executes inside an isolated sandbox — the AioSandboxProvider gives it a real filesystem and bash terminal in a Docker container (local, Docker, or Kubernetes-provisioned), while the LocalSandboxProvider is explicitly gated as not a secure boundary. Capabilities are extended through Agent Skills (SKILL.md capability modules, public or custom, loaded progressively) and through tools that include web search, web fetch, file operations, and bash, plus MCP servers exposed as deferred tools the agent must look up via a tool_search tool before calling. A persistent cross-session memory subsystem accumulates the user's profile and preferences, and a message gateway fronts the harness with LangGraph-compatible HTTP routes. The lead agent itself runs as a LangChain agent wrapped in a middleware chain (loop detection, summarisation, todo tracking, subagent limiting, sandbox auditing, and more). The original deep-research graph is catalogued separately as DeerFlow 1.x and is maintained on the `main-1.x` branch.

Agent loop shape. A lead agent runs a LangChain agent loop wrapped in a middleware chain; on each step it may call the `task` tool to delegate work to subagents (general-purpose, bash, or custom skill-scoped), each of which runs its own bounded ReAct-style loop inside an isolated Docker/AIO sandbox and returns a structured result. Concurrent subagent spawns are truncated to a clamped maximum, skills and deferred MCP tools are loaded lazily, and the lead agent synthesises subagent results into the final output while a cross-session memory subsystem persists profile and preferences across runs.