Large Language Models (LLMs) are increasingly used as cognitive interfaces that shape how users interpret evidence, communicate, and make decisions. As deployed systems become multimodal, tool‑using, and multi‑agent, safety failures no longer stem solely from isolated harmful outputs; they also arise from how agents update, store, and propagate epistemic commitments across retrieval, memory, tools, and peer communication.
In this work we treat belief as an engineering abstraction for evidence‑based commitments that guide downstream actions. Building on this view, we introduce the Explicit Belief Multi‑Agent System (EB‑MAS):
- Externalizes belief states onto a structured substrate;
- Employs belief operators to propose candidate updates;
- Applies a Cognitive Firewall to bound or reject updates that are weakly evidenced yet strongly driven by vulnerability cues.
We also examine risks at the group belief level, such as manufactured consensus, collusion, and cascading misinformation, and outline practical integration challenges:
- Grounding: aligning externalized beliefs with real‑world evidence;
- False positives: controlling the firewall’s tendency to block legitimate updates;
- Scalability: managing computational overhead for real‑time auditing in large agent networks;
- Ethical misuse: preventing adversaries from exploiting belief interfaces for manipulation.
Consequently, scalable agent safety requires more than better output filtering or larger base models; it demands explicit interfaces for tracing, calibrating, and constraining belief revision at runtime.
Review