Skill Blocks: How Should an Agent Load Its Skill? A Caching-Correct Comparison of Pre-load, On-Demand Tool-Loading, Progressive Disclosure, and Hybrid
This paper evaluates four skill-loading strategies for AI agents and demonstrates that while no single method is universally superior, hybrid and progressive disclosure approaches significantly reduce token consumption—particularly for large, multi-turn tasks—without compromising output quality.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the evolving world of artificial intelligence, there is a growing need to teach computer programs how to act like specialists. Instead of rewriting the core brain of the software, engineers now attach "skills"—detailed instruction manuals that tell the system how to perform specific tasks, such as searching for information, manipulating spreadsheets, or navigating a virtual environment. These manuals are read by the AI every time it is asked to do something. However, a practical problem has emerged: these manuals are often very long, containing hundreds of pages of rules and examples. When an AI is asked a simple question, it often has to read the entire massive manual, even if 90 percent of the text is completely irrelevant to the task at hand. This wastes computing resources and slows down the process, much like a librarian who insists on reading every book in the library before helping a customer find a single specific fact.
Researchers at Microsoft set out to solve this efficiency problem by asking a simple question: if you have a large skill manual, when should you show each part of it to the AI? They tested four different ways of delivering this information. The first method was the traditional approach: loading the entire manual every single time. The second method, called "Skill Block," kept only a tiny core of instructions visible and allowed the AI to ask for specific extra sections only when it felt it needed them. The third method, "Reference," showed the AI a short catalog of the manual's contents and let it pick which section to read next. The fourth method, "Hybrid," offered a very brief summary of every section upfront, allowing the AI to decide immediately if it needed the full details or if the summary was enough. The team tested these methods across five different types of tasks, ranging from single questions to complex, multi-step problems that required the AI to remember information over several turns of conversation.
The results revealed that there is no single best way to load these skills; the right choice depends entirely on the size of the manual and the nature of the task. For simple, one-off questions, the "Hybrid" approach proved to be the most efficient. By providing a short summary of every section, the AI could usually find what it needed without ever asking for the full text, saving a significant amount of data. In one test involving a large set of questions, this method reduced the amount of information the AI had to process by nearly 28 percent compared to reading the whole manual. However, for very small manuals or tasks where the AI needs the same instructions repeatedly, the extra effort of asking for information or reading summaries actually made the process slower and more expensive, offering no real benefit over just reading the whole thing at the start.
The findings became even more interesting when the researchers looked at complex, multi-step tasks where the AI has to remember context over a long conversation. In these scenarios, the "Skill Block" and "Hybrid" methods shined when the manuals were large and contained many sections that were rarely used. Because the AI only loaded the specific parts it needed for each step, it avoided the heavy cost of constantly resending the entire unused manual. In one complex simulation involving scientific reasoning, these on-demand methods reduced the effective data load by more than 60 percent compared to the traditional method. The researchers noted that this saving was so large that it outweighed the small cost of asking for the information. Conversely, in tasks where the instructions were short and needed for every single step, the advantage of these smart loading methods disappeared, and the systems performed roughly the same as the traditional approach.
A crucial part of this study was how the researchers measured cost. They realized that simply counting the total number of words sent to the AI could be misleading, especially in long conversations. Modern systems often remember previous parts of the conversation, so they do not need to fully re-process words that have already been seen. The team developed a way to count only the new information while giving a small discount to the repeated parts. When they applied this more accurate accounting, the benefits of the smart loading methods became even clearer for large, complex tasks. However, they also found that if the instructions were too small or if the AI had to ask for information too often, the overhead of managing those requests could cancel out the savings.
Ultimately, the study suggests that the best strategy for teaching AI skills is to match the delivery method to the job. If the manual is short or the task requires the same instructions constantly, it is better to just load everything at once. But if the manual is long and the AI only needs a small piece of it at any given time, a smart system that loads information only when needed is far superior. The researchers found that a "Hybrid" approach, which offers a quick overview of all options, works best for simple tasks, while a "Skill Block" approach, which lets the AI dig deeper only when necessary, works best for complex, long-running projects. This work does not claim to have found a universal rule that applies to every situation, but it provides a clear map for engineers to decide how to structure their AI's knowledge base to save time and resources without sacrificing performance.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.