SWE-Prime: Fewer Trajectories for Better Software Engineering Agents

SWE-Prime demonstrates that selectively curating a small set of high-quality training trajectories yields better software engineering agent performance than scaling data volume indiscriminately. The method achieves this by focusing on trajectory quality and relevance, not quantity.

This signals a potential deceleration in the compute-for-data arms race for agent training. For builders, the immediate operational shift is in data pipeline design: the bottleneck moves from accumulating diverse examples to developing robust filtering and scoring mechanisms for existing trajectories. This reduces the cost of dataset acquisition and curation, which is often a hidden tax on agent development. The second-order effect is on inference efficiency; fewer, better-conditioned examples may lead to leaner model architectures or reduced context windows for in-context learning, lowering per-task serving costs. Workflows centered on exhaustive synthetic data generation become less critical, replaced by targeted trajectory mining from production logs. Operators should audit their data ingestion pipelines for redundancy and prioritize investment in trajectory evaluation metrics over raw storage and generation capacity.