A tool error is either a message or a crash.
Larkspur Cloud’s support agent has two tools: search_handbook over the twelve handbook chunks from M10 (real TF-IDF) and get_server_status. Pick a question and a handle_tool_errors setting for the ToolNode: the trace shows every message the loop adds, and where a run stops.
- Messages in state
- Tool calls
- Errors the model read
- Outcome
The message trace
State after the run
| # | Message | Content |
|---|
Tool results and error texts are what LangGraph 1.2 produces for these calls, checked in Python. The model’s turns are scripted and representative: a real model chooses its own words and arguments.
All four questions at this setting
| Question | Outcome | Messages | Errors read |
|---|---|---|---|
| Answered | |||
A crashed run returns nothing to the customer: graph.invoke raises, and the messages so far are lost unless you checkpoint them.