When Agents Review Agents: Distributed Threshold Capacity and the Risk of Multi-Agent Laundering

Human oversight in AI governance is often assessed at the point where a person, review layer, or safeguard appears. This article argues that such assessment becomes insufficient in distributed and agentic systems, where action paths may be shaped across tools, agents, data pipelines, permissions, routing rules, and institutional handoffs before any visible reviewer arrives. I call the relevant capacity **Distributed Threshold Capacity**: the capacity of a socio-technical decision chain to preserve usable alterability—and the information required to recognize alterability—across multiple actors, systems, tools, and handoff points before an action path becomes materially, institutionally, or technically difficult to reverse. When this capacity is lost while each layer appears to perform review, the result is **multi-agent laundering**. The article’s central diagnostic claim is that local compliance can coexist with global capture: each layer may be procedurally correct while no layer retains an ordinary, authority-bearing route to reopen the path before binding. Through an illustrative composite scenario involving agentic benefits triage, the article shows how classification, scoring, flagging, routing, summarization, automated review, and late human approval can preserve the appearance of oversight while exhausting meaningful alteration upstream. The article further identifies **epistemic narrowing**, **silent binding**, and a **negative DTC standard** as diagnostic tools for detecting the loss of alterability. The contribution is diagnostic: it gives a vocabulary for locating the loss of pre-binding alterability without turning DTC into a technical control stack, legal liability theory, or universal audit protocol.