Context Injection Can Harm Multi-Agent Design Exploration, Study Finds
A recent study questions the belief that adding more context enhances multi-agent AI systems. Researchers evaluated seven conditions of context injection across ten software design tasks, conducting over 2,700 trials. They found a crossover effect: while certain artifact types increased design exploration on some tasks (up to 20× tradeoff coverage), they hindered it on others (up to 46% reduction). Interestingly, irrelevant documents sometimes matched or outperformed relevant ones. The influence of this effect is linked to baseline exploration without context, showing a Pearson correlation of -0.82 (p < 0.001). By adjusting convergence pressure through prompt design, two distinct regimes emerged. The authors suggest that context injection should be applied selectively. The study is accessible on arXiv, ID 2605.04361.
Key facts
- Study tests assumption that more context is better for multi-agent orchestration.
- Seven context-injection conditions tested across ten software design tasks.
- Over 2,700 runs conducted.
- Crossover effect found: same artifact type improves exploration on some tasks and degrades it on others.
- Improvement up to 20× tradeoff coverage; degradation up to 46% reduction.
- Irrelevant documents sometimes perform as well as or better than relevant ones.
- Direction predicted by baseline exploration without context (Pearson r = -0.82, p < 0.001).
- Two convergence regimes identified: natural (training data priors) and induced (explicit instructions).
- Natural convergence responds to artifact disruption; induced does not.
- Authors recommend conditional context injection.
Entities
Institutions
- arXiv