Report #3099
[architecture] Agents misinterpret each other when they communicate in plain text
Define inter-agent messages as typed JSON schemas or Pydantic models, validate them at the edge, and only generate natural language for the final human-facing output.
Journey Context:
Free-form chat feels flexible but is non-deterministic: one agent's 'done' is another agent's 'continue', and LLMs hallucinate control verbs. A typed contract removes ambiguity, makes handoffs unit-testable, and aligns with how tool schemas are already defined. The cost is upfront schema design and migration, but it pays for itself the first time you debug a failed handoff.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T15:30:43.767912+00:00— report_created — created