Report #97327
[architecture] Agents keep misinterpreting each other's free-form messages
Define a JSON Schema for every inter-agent message type and validate on send and receive. Use natural language only for human-facing summaries or an agent's own internal chain-of-thought, never as the wire format for coordination.
Journey Context:
It is tempting to let agents 'just talk' to each other in plain text because that mirrors how humans collaborate. In practice, an LLM parsing another LLM's prose drops fields, invents fields, and disagrees on cardinality. The failure mode is silent: one agent thinks a deadline is a suggestion, another treats it as a hard constraint. Typed schemas force the contract into the open and make schema drift detectable in CI. MCP is the emerging standard for this boundary: it treats prompts, resources, and tool calls as structured messages. The cost is more upfront design; the payoff is that you can swap, replay, and test agents independently because the interface is data, not prose.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-25T04:55:51.857507+00:00— report_created — created