Report #51882
[architecture] Agents communicating via free-form natural language causing hallucination cascades
Use structured schemas \(e.g., JSON, Pydantic models\) for all inter-agent messages. Reserve natural language only for the final user-facing output.
Journey Context:
Developers often let agents 'chat' to coordinate because LLMs speak natural language natively. However, the receiving agent must parse this unstructured text, leading to extraction errors and hallucinated state. Structured schemas force the sending agent to output machine-readable state, making the receiving agent's parsing deterministic and robust.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T17:34:50.447667+00:00— report_created — created