We Replaced Message Buses with Telemetry for AI Agent Coordination

The article discusses a new framework called BossCat Protocol that uses OpenTelemetry as the primary coordination mechanism for AI agents, instead of traditional message buses. This approach has improved debugging, scalability, and reduced coordination overhead.

Details

The article explains that the authors have been working on building multi-agent systems in production for 2.85 years. They found that the typical approach of using message buses for agent coordination creates several challenges, such as coordination overhead, debugging nightmares, scaling issues, and complex state management. To address these problems, they developed the BossCat Protocol, which uses OpenTelemetry as the primary coordination mechanism. In this approach, agents emit structured telemetry to a shared observability backend, and they can then query this telemetry to understand what they and other agents have done. This creates an 'emergent coordination' where agents naturally coordinate based on the shared source of truth, without the need for explicit message passing. The authors also built an 'evidence-first governance' framework on top of the telemetry-based coordination, where agents must provide specific telemetry query results as proof of task completion before proceeding. This has led to a 96% quality gate pass rate and significant improvements in debugging and infrastructure complexity.