LangChain Deep Agents 0.10.5 Returns Tool Errors to the Model Instead of Failing the Run

LangChain · x · 2026-11-20

LangChain's caspar shares a small reliability win in Managed Deep Agents 0.10.5: instead of terminating the run on tool errors, most errors are returned to the model with a status flag so it can recover and retry.

Most tool errors are recoverable, so this change makes agents noticeably more resilient for long-running work. The accompanying diagram shows the before/after: previously the agent ended with "An error occurred"; now the model sees the error, adjusts, and re-executes successfully.

Original post →

More from coding & agent

coding & agent channel →