// HASTE: tiered skills for ML engineering agents // Why do ML engineering agents keep rediscovering the same techniques on every new task? …

// HASTE: tiered skills for ML engineering agents // Why do ML engineering agents keep rediscovering the same techniques on every new task? Because each competition is a cold start. New research introduces HASTE, a hierarchical multi-agent system that organizes accumulated skills into three scope tiers, global, domain, and competition-specific, each bound to a matching agent level. An orchestrator coordinates domain specialists and promotes learning upward through LLM-driven abstraction. The ablation is the headline. Holding a 159-skill inventory fixed across 8 competitions, tiered loading reaches a 100% medal rate while flat loading reaches 62.5%, the same rate as loading no skills at all, and burns twice the output tokens. On MLE-Bench Lite it hits a 77.3% medal rate across 22 Kaggle competitions. Why does it matter? How you scope memory can matter more than how much you accumulate, and better knowledge organization can partly substitute for model strength and compute. Paper: https://arxiv.org/abs/2606.30911 Learn to build effective AI agents in our academy: https://academy.dair.ai/

Source: DAIR.AI (X) | 2026-07-05