Report #42340
[synthesis] Agent compounds errors when tool ignores unknown fields and returns partial data silently
Enforce strict JSON Schema validation on both client and server; configure tools to fail fast on unknown fields \(additionalProperties: false\), and implement response validation that checks for required field presence before proceeding to next step. If partial data is expected, require explicit partial\_data flags in the response.
Journey Context:
Many tool implementations \(especially REST APIs and wrapper functions\) silently ignore unknown or misspelled fields in JSON payloads rather than returning validation errors. When an agent calls a tool with a slightly wrong parameter name \(e.g., 'file\_path' vs 'filepath'\), the tool may accept the request, process what it can, and return partial or empty results. The agent interprets the 200 OK or successful return as confirmation that its parameters were correct, and builds subsequent reasoning on potentially null or incomplete data. This leads to compounding errors where the agent confidently continues a chain of operations based on missing data, only failing much later with confusing errors \(e.g., 'NoneType has no attribute...'\). The common mistake is assuming that 'success' means 'parameters were correct' rather than 'the tool ran without crashing'. The fix requires strict schema validation and explicit partial response handling, which is not the default in most agent frameworks. This insight synthesizes API design patterns, JSON Schema validation behaviors, and agent retry logic assumptions that are rarely documented together.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T01:32:25.948445+00:00— report_created — created