A Config-Driven Control Plane for Human-in-the-Loop Multi-Agent Systems

This piece outlines a production-shaped reference architecture for running many concurrent AI agents under human supervision, arguing that the hard problem is coordination and authority, not model quality. It reframes the interface as a forum of workstreams and threads rather than a chat window, letting one operator track many parallel agent sessions from a single view.

The core design freezes a small generic engine (Watcher, Synthesizer, Hub) on top of a shared pub/sub fabric, routing work by capability_id instead of agent address so new agents plug in through configuration and topics rather than new integration code. A fixed three-message protocol - HQ (query, blocking), HR (authoritative response), and HI (non-blocking status) - standardizes every human-agent interaction across the system.

The architecture closes the full loop - detect, propose, approve, execute, record, learn - backed by an append-only store for audit and downstream learning, with routing and registry components kept stateless and horizontally scalable. For engineers, the takeaway is that scaling an agent fleet is fundamentally a routing and protocol problem: sub-linear growth in engineering effort comes from freezing the core engine and pushing all domain logic into declarative capability catalogs and pluggable executors.