The study investigates the effectiveness of different approaches to leveraging reusable knowledge within language model agents. Recent methods utilize agent skills, represented as skill packages, to guide agent behavior. These packages are typically loaded into the main agent’s context window, relying on the agent to follow the instructions. As task horizons increase, this method becomes less reliable due to the accumulation of information within the context window, leading to degraded reasoning quality. The research proposes an alternative: invoking skill packages as separate subagents. This approach creates dedicated context windows for each subtask, isolating the reasoning process.
The results indicate that subagent execution outperforms agent-skill execution when skill packages have clear input-output contracts and their instructions encode procedural knowledge. This suggests a structured approach to skill design is critical for optimal performance. The study highlights the importance of organized knowledge representation and invocation strategies. The core finding is that the benefit of reusable knowledge depends on both its content and how it is organized and invoked.
Communication overhead is identified as a tradeoff. Extra tokens are required to coordinate between the main agent and its subagents. This overhead needs to be considered when evaluating the overall efficiency of the subagent system. The research suggests that the performance gains from using subagents outweigh the communication costs for complex, long-horizon tasks.
Source: https://arxiv.org/abs/2609.09233