Researchers have a fix for a costly problem in AI systems built from multiple cooperating agents: watching every agent all the time is expensive, but watching only at the end lets errors spread too far before anyone notices.
The new method, called BRA-Audit, treats a multi-agent AI system as a dependency graph and decides where to place a limited number of "audit checks" to catch problems as they happen. Instead of auditing everything or waiting until the end, a greedy scheduler targets the most influential and longest-unchecked parts of the workflow. When it flags an issue, the system can roll back just that piece instead of the whole run. Tested across coordination, reasoning, and open-ended tasks, it performed close to a fully audited baseline while cutting total token usage by 17.2% to 40.6%.
This matters because multi-agent AI setups are becoming standard for complex work, and a single hallucinating or compromised agent can quietly corrupt everything downstream. Full-coverage auditing catches that but burns tokens fast enough to erase the cost savings of using multiple cheaper agents in the first place. A scheduler that spends its audit budget where damage would spread furthest is a more realistic answer than either extreme.
It is also a tacit admission that current multi-agent AI deployments are running with too little oversight, too late, because nobody wants the bill for auditing everything.