Subagents Outperform Agent Skills for Some Long-Horizon Tasks

Summary

The paper studies how language-model agents can use reusable knowledge packages during long-horizon tasks. Conventional agent-skill execution loads a package's instructions into the main context and asks the agent to follow them, but this can become brittle as accumulated context affects reasoning quality. The authors instead invoke each package as a subagent with a fresh context window dedicated to a subtask. Subagent execution performs better when a skill package has a clear input-output contract and instructions that encode the procedure needed to satisfy it. This approach requires extra tokens for coordination between the main agent and its subagents. The results suggest that the usefulness of reusable knowledge depends not only on what the package contains, but also on how it is organized and invoked.