Byzantine-Resilient and Human-Observable Agentic Systems for Scientific Discovery

We argue an AgenticOS must treat agent outputs as proposals for state transitions, grounded in evidence and provenance, rather than as truth. We introduce four coordination primitives: propose, validate, commit, and escalate, through which Byzantine-resilient consensus authorizes state changes, while unresolved disagreement is exposed as a meaningful system state rather than a failure. This reframes non-consensus as a natural entry point for human scientific judgment, and raises questions about defining Byzantine failure for probabilistic, tool-using agents.