Subagents Outperform Direct Skill Loading in Long-Horizon AI Tasks

A recent preprint on arXiv explores how language model agents can better leverage libraries of reusable knowledge when executing long-horizon tasks. Currently, agents typically load skill packages—bundles containing instructions and scripts—directly into their primary context window. However, reasoning capability degrades as context accumulates over extended tasks.

To address this limitation, researchers investigated an alternative architecture that invokes skill packages as distinct subagents. Instead of crowding the main context, this method spawns fresh context windows dedicated to specific subtasks. The findings show that subagent execution consistently outperforms standard skill execution when skill packages feature clear input-output contracts and structured procedural instructions.

The main trade-off is an increase in communication overhead, requiring additional tokens to coordinate between the primary agent and its subagents. The authors conclude that the utility of reusable knowledge relies heavily on how it is architected and executed.