Subagents vs. Agent Skills: Executing Reusable Knowledge for Long-Horizon Agentic Tasks
arXiv:2609.09233 [cs.AI] — Submitted 7 Sep 2026
Authors: Wasu Top Piriyakulkij, Rachel Lawrence, Alicia Curth, Sushrut Karmalkar, Niranjani Prasad
Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)
Abstract
How can language model agents effectively leverage libraries of reusable knowledge to solve long-horizon tasks?
Recent work has increasingly focused on agent skills: reusable capabilities represented as skill packages — multi-file bundles containing instructions, scripts, and other resources that help agents perform specific tasks. Agent skills are typically executed by loading their instructions into an agent's context and relying on the agent to follow them. As task horizons grow, however, this approach becomes increasingly brittle, because reasoning quality degrades as more information accumulates in the context window.
This paper investigates an alternative approach in which skill packages are instead invoked as subagents. Rather than loading skill instructions into the main context, subagent execution spawns fresh context windows dedicated to solving individual subtasks.
The authors show that subagent execution outperforms agent-skill execution when skill packages expose clear input–output contracts and their instructions encode the procedural knowledge needed to fulfill those contracts. The tradeoff is additional communication overhead, as extra tokens are required to coordinate between the main agent and its subagents.
The results suggest that the benefit of reusable knowledge depends not only on its content, but also on how it is organized and invoked.
Why This Matters in 2026
As agentic systems move from short, single-turn prompts toward long-horizon, multi-step workflows, context management has become a central bottleneck. The 2026 research landscape reflects this shift: rather than treating skills as passive documentation, leading approaches treat them as invocable units — much like function calls in a program — with well-defined interfaces and isolated execution contexts.
This paper's framing sits at the intersection of two emerging trends:
- Modular skill packaging, where capabilities ship as portable, multi-file bundles that agents can compose.
- Context isolation via subagents, where each subtask runs in a clean context window to preserve reasoning fidelity.
- Skill packages bundle instructions, scripts, and resources into reusable capabilities for LLM agents.
- Agent-skill execution loads instructions into the main context and relies on the agent to follow them — increasingly brittle as task horizons grow.
- Subagent execution spawns fresh context windows per subtask, preserving reasoning quality.
- Subagents win when skills have clear input–output contracts and encode procedural knowledge.
- The cost: extra communication tokens between the main agent and its subagents.
The practical implication is that architectural choices — not just knowledge content — determine whether reusable skills pay off. Teams designing agent systems in 2026 should weigh the coordination cost of subagent invocation against the degradation cost of a polluted main context.
Key Takeaways
Bottom line: The value of reusable knowledge in agentic systems depends not only on what is stored, but on how it is organized and invoked.
Cite as: arXiv:2609.09233 cs.AI] · DOI: [10.48550/arXiv.2609.09233
via ArXiv AI