Giving an AI agent a stack of instructions to follow is not the same as giving it a coworker to hand the job to.
A new arXiv paper compares two ways of running "agent skills" - reusable packages of instructions, scripts, and resources meant to help language model agents handle specific jobs. The standard approach loads a skill's instructions straight into the agent's main context window and trusts the agent to follow them step by step. The researchers tested an alternative: treating each skill package as a subagent, spawning a fresh context window to handle that subtask instead of piling more text into the original one. Across long-horizon tasks, the subagent approach won, but only when a skill package had a clear input-output contract and instructions that actually encoded the procedure needed to meet it.
The finding lands on a real pain point. Context windows are not infinite, and reasoning quality drops as more instructions, scratch notes, and tool output accumulate in them - which is exactly what happens when an agent tries to juggle a dozen subtasks without a change of scenery. Splitting work into subagents fixes that by giving each subtask a clean slate, at the cost of extra tokens spent coordinating between the main agent and its helpers.
That tradeoff is the real story: subagents are not free, they are a bet that clean context beats efficient token use. As more products ship "skills" as their headline feature, this is a reminder that how reusable knowledge is packaged matters as much as what it says.