AgentFuse: Predicting Cascading Failures in Multi-Agent LLM Pipelines
Middleware that predicts where multi-agent pipelines will break — before they break.
What this is
Most multi-agent LLM pipelines fail in the seams: an upstream agent’s plausible-but-wrong output gets passed to a downstream agent that confidently builds on it, and the error cascades. AgentFuse is a middleware layer that sits between agents in a pipeline and predicts the likelihood that any given hand-off is about to propagate a failure downstream.
The contribution is methodological as well as empirical: the framework treats inter-agent hand-offs as the primary unit of analysis, rather than individual agent outputs. This reframing turns out to be more predictive of end-to-end pipeline failure than scoring agents individually.
Why it matters
The pattern of cascading failure in agent pipelines is currently invisible to standard evaluation. Per-agent benchmarks miss it by construction. Without an instrument for the hand-off layer, the field has no good answer to “where will this pipeline fail in production?” — which is the question deployment teams actually ask.