Multi-agent harnesses need separate message channels

Current multi-agent setups funnel every message into one chat, mixing supervisor directives, worker reports, co‑worker edits, peer coordination, and system alerts into a single scroll. That makes it impossible for a human supervisor to quickly tell who said what, who has authority, or which items need action. Five distinct relationships matter: upstream (the supervisor setting goals and scope), downstream (agents executing tasks), sideways (a human co‑worker pairing on the same work), peers (other agents coordinating laterally), and alerts (automated events like CI failures). Upstream and sideways often look similar but carry different authority and expectations, and simple tagging or prefixes can't restore readability or auditability once streams are interleaved.

A proper harness should enforce those five channels inside one shared context: tag messages at the protocol level, present separate panes or filters for each channel, require the agent to choose the reply target, and encode authority and routing rules so upstream can decide "done," sideways can override implementation, downstream supplies input, peers coordinate, and alerts follow severity and deduplication rules. That approach preserves clear audit trails, prevents peer collisions and misdirected replies, lets humans step in cleanly, and keeps models from having to infer social roles. As single-agent tooling is chained into workflows, chat-first UIs need channel structure or they won't scale.