This post summarizes a research paper presented at the ACM CAIS conference on simulation-driven resilience in agentic data systems. It begins with metastability — the failure mode where protective mechanisms like retries and queues become amplifiers that keep systems perpetually behind. AI agents worsen this by retrying aggressively, fanning out into bursty parallel tasks, and holding transactions open while waiting on LLMs, creating ~20x more branches and ~50x more rollbacks than humans. The authors propose MESSI (MEtaStability SImulator), a discrete-event simulation framework that models distributed systems as directed graphs to explore and detect metastability before production incidents. Two key findings from simulating an Execution Control System (ECS): first, two individually sensible queue-probing policies can compose into a metastable feedback loop where ticket inflation causes performance collapse; second, naive admission control can destructively interfere with the ECS by preventing it from rebalancing under load, leaving the system stuck in reduced-throughput mode — a problem amplified by agents whose retries are indistinguishable from first attempts.
