When Several AI Agents Work Together: How Firm Is the Monitor's Verdict? One Warning Number for Each Agent

Recent papers on watching AI agents each end in a single verdict from a monitor: a suspicion score, a yes or no, or an alarm that stops the run. None of them says how firm that verdict is. This concept note shows the open gap in five of these papers (arXiv 2610.24052, 2610.02994, 2605.08715, 2605.31593 and 2607.24893). The idea is simple. Read the same monitor several ways, using only the settings it already permits: the normal setting and settings near both ends of the permitted range. The same program is run, unchanged, on each version. The spread between the versions is inversely proportional to the size of the leading eigenvalue, so a wide spread means the verdict is near a tipping point. One series is kept for each agent. Several agents working together are therefore several separate series, and holding back any one agent breaks a harmful chain. A rising warning number makes the computer decide a shunting action, without waiting for a person. The action moves load aside, for example by lowering the agent's priority or sending its actions through a closer check, and it does not stop the work. The note also lists first tests that can be run on data the authors of these papers already hold. Patent pending. Keywords: AI agents, multi-agent systems, monitoring, early warning, loss of control, distributed attacks, warning number, shunting action