Agent pairs fail together 90% of the time, study finds

A new arXiv paper tested 18,000 multi-agent missions and found that when one agent in a pair fails, the other fails too 90.0% of the time, per arXiv 2608.12895 (2026-08-13). The log odds ratio is 6.66, with a 95% confidence interval of 6.38 to 7.00. That's not noise. That's agents failing together, not independently.

Here's why that number matters. The standard way to size a multi-agent system is to multiply per-agent success rates. Two agents at 95% reliability each, run independently, should give you a combined failure rate under 1%. That math only works if the failures are actually independent. This study says they aren't. Failures share a cause, whether that's a bad prompt context, a shared tool outage, or a correlated blind spot in the underlying model, and the paper's preregistered design means this wasn't a post-hoc pattern someone went looking for.

The practical hit lands on anyone who's added a second agent, a verifier step, or a redundant check and called it a reliability win. If your two agents share an upstream data source, a model family, or a failure-prone tool call, you likely haven't bought the safety margin the arithmetic implies. The paper doesn't say how much margin you actually get back, only that treating agents as independent overstates it.

What I'd want to see next: whether the 90% co-failure rate holds when the two agents genuinely don't share infrastructure, meaning different model providers, different tool sets, different prompts written by different people. If co-failure drops once shared dependencies are removed, that tells you where the correlation actually lives, and it's a cheaper fix than redesigning your reliability math from scratch.