The Observability Gap in AI Multi-agent Systems

Abstract

As AI agents become more widely deployed, take on more tasks and help complete work projects, there will be an increase in interactions between them, leading to failures that are particular to multi-agent systems, far beyond those of deployments of single agents in isolation. Real-time monitoring is necessary to prevent these failures because both pre-deployment testing, while necessary, is insufficient, and post-hoc auditing does not prevent high-stakes incidents. However, agent frameworks have significant gaps in capturing key telemetry signals for effective real-time monitoring. This paper provides an empirical assessment of these gaps and provides preliminary recommendations for addressing them.