Equipping large language model agents with reusable domain knowledge has increasingly centered on agent skills—structured bundles of instructions and scripts designed for specific tasks. A research team comprising Wasu Top Piriyakulkij, Rachel Lawrence, Alicia Curth, Sushrut Karmalkar, and Niranjani Prasad from Cornell University and Microsoft Research Cambridge evaluated how language models execute these knowledge libraries across extended workflows.
The Context Bottleneck in Skill Execution
Under the standard paradigm, agent skills operate by loading their instructions directly into an agent's main context window and relying on the model to follow them. As task horizons expand, this mechanism encounters severe structural limits. Reasoning quality steadily degrades as instructions and intermediate outputs accumulate inside a single context window, rendering long-horizon workflows brittle.
"As task horizons grow, however, this approach becomes increasingly brittle, because reasoning quality degrades as more information accumulates in the context window."
Each agent skill is defined as a skill package: a multi-file package containing instructions and scripts that help solve specific tasks. When an agent continually appends these instruction sets into its active context, the accumulating information strains the underlying model's capacity to maintain coherent reasoning across multi-step objectives.
Delegating Execution to Isolated Subagents
To bypass context bloat, the researchers investigated an alternative architecture: invoking skill packages as dedicated subagents. Rather than loading instructions into the primary context, subagent execution spawns fresh context windows dedicated to solving individual subtasks. The subagent performs the required work independently within its clean workspace and returns only the final output to the main controller.
The researchers evaluated procedural, contract-driven skill packages. Subagent execution systematically outperforms in-context skill execution when skill packages expose clear input-output contracts and their instructions encode the procedural knowledge needed to fulfill those contracts. By partitioning operations across distinct environments, the system reduces the maximum volume of information processed by any single context window.
Coordination Tradeoffs in Modular Architectures
The structural advantage of subagent modularity introduces a direct operational tradeoff in the form of communication overhead. Extra tokens are required to coordinate and pass messages between the main orchestrator and its delegated subagents. The utility of reusable knowledge repositories in autonomous pipelines depends not merely on the quality of their instructional content, but fundamentally on whether that knowledge is organized and invoked through isolated execution environments.