ToolNode and tools_condition

ToolNode executes tool calls from the last AI message. tools_condition routes to that node when tool calls exist, or to END when they do not.¹

Tiny example

from langgraph.graph import START, StateGraph, MessagesState
from langgraph.prebuilt import ToolNode, tools_condition

builder = StateGraph(MessagesState)
builder.add_node("model", call_model)
builder.add_node("tools", ToolNode([lookup_order]))
builder.add_edge(START, "model")
builder.add_conditional_edges("model", tools_condition)
builder.add_edge("tools", "model")

The loop is: model proposes a call → tools execute → model reads the result → model calls again or answers.

The default error boundary

Event inside `ToolNode`	Default behavior
Model supplies invalid tool arguments	Return an error `ToolMessage` so the model can correct the call
Tool function raises during execution	Re-raise the exception so the graph can fail or retry
`handle_tool_errors=True`	Catch all tool exceptions and return error `ToolMessage` values

Failure note

If handle_tool_errors=True catches a transient execution error, the node’s RetryPolicy never sees that exception. Choose deliberately: error message for model correction, or escaped exception for system retry. ToolNode can execute multiple tool calls in parallel, so every write tool must tolerate a repeated node attempt.

LangChain, Tools — ToolNode and ToolNode API reference. ↩