Production agents interact with real APIs that return errors, empty results, malformed JSON, and rate limit responses. Building a robust tool result parsing layer — with typed schemas, error classification, retry logic, and graceful degradation — is what separates reliable agents from brittle demos.

In production, tools fail constantly. APIs return 429s. Search results are empty. Database queries time out. File reads return binary instead of text. Web scrapes return JavaScript-heavy pages that render as near-empty HTML. A fragile agent passes all of this raw to the model and produces confused, incorrect, or stuck behavior. A robust parsing layer handles it systematically. Error classification is the first step: distinguish between retriable errors (rate limits, timeouts — retry with backoff), empty results (the query had no matches — signal clearly to the agent so it can rephrase or try a different tool), data format errors (malformed response — log and fall back to a secondary tool), and hard failures (authentication error, endpoint removed — escalate to the operator immediately). Schema enforcement using Pydantic or JSON Schema validation on tool outputs ensures that agents receive data in the expected format and get explicit error signals when they don't, rather than silently processing corrupted data. Result truncation must be content-aware: truncating a JSON object mid-key produces unparseable garbage; truncation should happen at logical boundaries (whole records, whole paragraphs). For tools that return large structured datasets, a secondary summarization step — using a fast, cheap model to extract the top N most relevant records for the current task — is more token-efficient than naive truncation. Logging and observability of the parsing layer is the most direct diagnostic tool for debugging agent failures in production.

Robust Tool Result Parsing: Error Handling and Schema Enforcement in Production Agents