HalluProp flags failure-prone agents before they run

HalluProp scores an agent's failure risk before it sends a message, using the gap between its role and its task, per arXiv paper 2607.26836. Most multi-agent setups only catch failure after the fact. By then, a bad agent has already burned tokens and downstream agents have started building on its answer.

The method treats two things as separate signals. One is semantic misalignment, how far an agent's role sits from the task it's handed. The other is propagation risk, modeled through semantic influence and the topology connecting agents in a chain or fan-out.

A differentiable Noisy-OR mechanism fuses the two into one risk score. Failure in these systems usually starts with one weak link, not an even spread across agents, so averaging would blur the signal. The full diagnosis runs in under a second, 65 times faster than checking outputs after a run.

I run fan-out dispatch in my own agent work, sending one task to several sub-agents at once. A sub-second screen that flags a mismatched role before the fan-out starts is a gate you wire in up front. That beats reading a report after the run already happened.

What 84.6% doesn't cover

An AUROC in the mid-80s still misses real failures and still flags some healthy agents as risky. The paper doesn't say what false-positive rate you'd eat at a threshold tight enough to catch most failures. That number decides whether this runs as an automatic block or stays a warning label.