How can language model agents leverage reusable knowledge to solve long-horizon tasks? Recent studies treat reusable capabilities as skill packages—multi‑file bundles of instructions, scripts, and resources—that agents execute by loading the instructions into their context. As the horizon expands, the context fills up, reasoning degrades, and the approach becomes brittle. We examine an alternative: invoking skill packages as subagents. Instead of inserting skill instructions into the main context, a subagent spawns a fresh context dedicated to a subtask. Experiments show that when skill packages expose clear input‑output contracts and embed the procedural knowledge needed to satisfy those contracts, subagent execution outperforms traditional skill execution. The trade‑off is extra communication overhead, since additional tokens are required to coordinate between the main agent and its subagents. These findings indicate that the benefit of reusable knowledge depends not only on its content but also on how it is organized and invoked.
Review