When Context Hurts: The Crossover Effect of Knowledge Transfer on Multi-Agent Design Exploration

A new arXiv preprint from Saranyan Vigraham challenges the prevailing wisdom that more context always improves AI agent performance. The study, titled "When Context Hurts: The Crossover Effect of Knowledge Transfer on Multi-Agent Design Exploration," ran over 2,700 multi-agent experiments across 10 software design tasks under 7 different context-injection conditions. The results reveal a clear crossover effect: giving agents context (e.g., design artifacts, documentation) boosted tradeoff coverage by up to 20x on some tasks, but actively degraded performance—up to 46%—on others. Strikingly, on several tasks, an irrelevant document performed as well as or better than every relevant artifact. The direction of the effect is predicted by a single measurable variable: baseline exploration without context, with a Pearson correlation of -0.82 (p < 0.001).

To probe the underlying mechanism, the author manipulated convergence pressure through prompt design. This revealed two distinct regimes: convergence driven by training data priors (natural convergence) responded to artifact disruption, while convergence driven by explicit instructions (induced convergence) did not. The key implication is that context injection should be conditional, not universal. A single no-context trial serves as a cheap diagnostic that predicts whether knowledge artifacts will help or hurt a given task. For engineers building multi-agent systems, this means they can quickly benchmark tasks before blindly injecting RAG documents or examples—turning a long-held assumption into a testable hypothesis.