Report #94542
[synthesis] Agent misinterprets structured tool output at type boundaries and compounds the error through subsequent steps
After every tool call, have the agent explicitly restate its interpretation of the output before acting on it. Use structured output formats \(JSON\) over unstructured text. Add a parse verification step that checks key fields exist with expected types. When output is ambiguous, halt and clarify rather than guessing.
Journey Context:
Agents systematically misinterpret tool output at type boundaries: reading a table column as a row, treating a nested JSON field as top-level, confusing a list-of-one-object with the object itself, or misreading multi-line string boundaries. The synthesis across three domains reveals why this compounds catastrophically: \(1\) NLP parsing research shows boundary errors are the most common and least detectable parsing failure; \(2\) agent tool-use patterns show agents rarely re-verify their interpretation before acting; \(3\) error propagation analysis shows boundary misinterpretations compound multiplicatively, not additively—each subsequent step transforms the wrong data, making the original error harder to trace. The key insight: a misinterpreted list-of-lists vs. a flat list doesn't produce an error in step 2. It produces wrong data that gets transformed in step 3, joined in step 4, aggregated in step 5. By step 6, the error is unrecognizable from the original misinterpretation. The 'restate before acting' pattern forces the agent to commit to an interpretation that can be verified, catching plausible-but-wrong readings before they compound.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T17:16:21.881917+00:00— report_created — created