How AI Agents Should Handle Tool Failures Without Crashing

When a tool fails inside an AI agent, propagating the exception the way traditional software does is counterproductive — the error must instead be caught, formatted clearly, and passed back to the model as a readable message. Well-structured error messages should explain what failed, why it failed, what is actually true, and what the model should do next, making that final directive the most valuable text in the entire tool layer. Raw tracebacks are wasteful because they consume hundreds of tokens and are re-sent on every subsequent request throughout a run, compounding costs significantly. Errors fall into distinct categories — transient, malformed-args, not-found, empty-result, forbidden, and corrupt-state — each requiring a different response, from silent code-level retries to immediate escalation. A particularly overlooked case is the empty-result, which is technically a successful call but frequently triggers infinite retry loops because a model interprets 'no results' as a near miss rather than a definitive answer.