Artificial intelligence agents depend on tools to affect systems and retrieve current information. A tool error can originate in the agent's selection or arguments, the tool contract, authentication, rate limits, network availability, application state, or the returned data. Clear error categories help separate these causes instead of treating every failure as a model mistake.
Good tool error handling returns bounded, machine-readable information and identifies whether retrying is safe. Agents should not repeat consequential operations blindly, expose credentials in logs, or interpret a permission denial as evidence that data is absent. Evaluation should track both error frequency and successful recovery.
