Error Recovery: What an Agent Should Do When a Tool Fails

Generic error messages like 'Something went wrong' are harmful for two-thirds of AI failures, as retrying is often futile. Errors originate from four layers, including transport issues and rate limits, requiring distinct responses. For AI agents, crashing during tool failure wastes their core capability. Instead of letting exceptions propagate, developers should catch, format, and append errors as tool messages. This allows the model to read the failure as information and attempt alternative actions.