Coordination and Communication Protocols for Multi-Agent Financial AI Systems in Automated Trading

The opacity of machine learning models in automated financial trading remains a fundamental challenge, further exacerbated in multi-agent systems by information decay and the lack of formal coordination mechanisms. More recent LLM-based models, including Sleipnir and TradingAgents, are highly coordinated with non-deterministic inference, but do not provide mechanisms to maintain the integrity of explanations, or systematically deal with concept drift. We developed a scalable multi-agent control protocol with a five-agent system where all communication is mediated by a typed and message bus, based on file-backed persistence. The framework presents three major contributions: (C1) a typed coordination protocol with SHA-256 checksums and validation gates to enforce artefact integrity; (C2) a buffered retraining that prevents noise-based updates by enforcing consecutive drift confirmation; and (C3) SHAP-based cross-run drift detection with feature importance consistency. Assessed on eight assets with five years (2019–2024) of operation DMACCP shows a reduction in unnecessary retraining of 66.7% with no statistically significant difference to a No-MAS control. The suggested framework shows that explainability-based coordination of real-world financial AI systems are possible and that scalable multi-agent systems can be built in automated trading and beyond.