'Smarter Saboteurs, Better Fixers': scaling security in linear multi-agent workflows

arXiv·medium signal

This paper (arXiv 2606.12709, submitted June 12) studies adversarial resilience in sequential multi-agent task pipelines, modeling how increasingly capable 'saboteur' agents and 'fixer' agents co-scale within linear workflows. It offers a framework for reasoning about where injected errors propagate and where verification agents should sit. For builders chaining agents, it formalizes the intuition that a verifier's value depends on the sophistication of the failure it must catch.