Source: LangChain multi-agent docs (overview and handoffs)

LangChain v1 docs comparing subagents, handoffs, skills and router patterns, and showing handoffs as state-updating tools that return Command.

Contents
  • What: official documentation pages (Python): multi-agent overview and the handoffs pattern
  • Publisher: LangChain
  • Date: undated living docs (read 2026-11-14)
  • Original: Multi-agent overview, Handoffs

Summary

The overview lays out five multi-agent patterns (subagents, handoffs, skills, router, custom workflow), who keeps control in each, and worked call/token counts. The handoffs page defines a handoff as a tool that updates a persisted state variable (active_agent, current_step), after which the system either reconfigures one agent or routes to another agent node.

Key claims

  • Handoffs: a tool call updates shared state and the active agent persists across turns; subagents: the main agent calls specialists as tools, they start fresh each call, and all routing returns to the main agent.
  • Worked example counts: one-shot request, subagents 4 model calls vs 3 for handoffs; repeating the request in turn 2, handoffs total 5 calls vs 8 for subagents; a three-domain request, handoffs 7+ calls and ~14K+ tokens vs subagents 5 calls and ~9K tokens.
  • Handoffs run sequentially and cannot research several domains in parallel; they rate highest for direct user interaction. Subagents rate highest for parallelism and distributed development, lowest for direct user interaction.
  • Two implementations: a single agent whose prompt and tools change by state via @wrap_model_call middleware (recommended for most cases), or separate agent subgraphs navigated with handoff tools returning Command(goto=..., update=..., graph=Command.PARENT).
  • Handoff tools read state through ToolRuntime; when they update messages, they must include a ToolMessage matching runtime.tool_call_id, or history is malformed.
  • Recommended context: pass only the AIMessage holding the handoff call plus its ToolMessage, not the whole subagent history, which confuses the receiver and costs tokens; summarise into the ToolMessage if more is needed.
  • The examples assume a single tool call per turn; parallel tool calls need separate handling. A turn handed back to the user should end on an AIMessage.

Feeds

Suggest changes →