SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries
SkillZip is an execution-aware framework that enables scalable agent skill libraries by performing contract-preserving, section-level graph compression to generate compact, dependency-closed contexts that maintain procedural integrity and outperform existing baselines in both efficiency and retrieval accuracy.
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
Imagine you are teaching a super-smart robot how to do chores, like cooking a meal or organizing a messy room. You can't just give the robot a giant, 500-page instruction manual for every single task; it would get overwhelmed, forget the important parts, or get stuck reading about how to bake a cake when you just asked it to fold laundry. This is the world of AI Agents: computer programs that use "Large Language Models" (the same kind of brain behind chatbots) to figure out how to use tools and solve problems. To help them, we give them a "skill library"—a collection of reusable instructions, like a recipe book or a toolbox.
But here's the tricky part: as these libraries grow to hold thousands of skills, they become messy. If you ask the robot to "clean a CSV file," it might pull up a whole book on "Data Analysis" that includes chapters on making charts and writing reports, even though it only needs the first three steps to clean the file. This wastes the robot's memory and confuses it. The big question scientists are asking is: How can we shrink these massive instruction libraries down to the tiniest, most useful pieces without losing the safety checks and rules that keep the robot from making mistakes?
This is where a new system called SkillZip comes in. Think of it as a magical compression tool for robot instruction manuals. Instead of treating a whole skill (like "Clean CSV") as one giant, unchangeable block, SkillZip breaks every skill down into its smallest working parts: the "Intent" (what we want to do), the "Operation" (the actual steps), and the "Verifier" (the safety check to make sure it worked).
The researchers found that many of these small parts are identical across different skills. For example, "Clean CSV" and "Pivot Table" both need to do the exact same thing to read a file and check if the rows are correct. Old systems would load the entire "Clean CSV" book and the entire "Pivot Table" book, wasting space. SkillZip, however, spots these repeating patterns and turns them into macros—like reusable stickers or macros in a spreadsheet. It says, "Hey, we've seen this 'read file' step a hundred times; let's just store it once as a tiny, reusable macro."
But there's a catch. If you just shrink text, you might accidentally delete a safety rule, like "make sure the file isn't empty before you start." That would be dangerous. SkillZip is special because it is contract-preserving. It acts like a careful editor who only shrinks the text if the "contract" (the rules, inputs, outputs, and safety checks) stays perfectly intact. It ensures that even though the instruction is now a tiny macro, the robot knows exactly what to feed it and how to check if it worked. If the robot needs more detail later, SkillZip can instantly "unzip" the macro back into the full, original instructions.
In their experiments, the team tested this on two different types of robot challenges: one involving data processing (like cleaning spreadsheets) and another involving a robot navigating a virtual house. They found that SkillZip was incredibly effective. It managed to shrink the skill library by 3.46 times (making it less than a third of its original size) while keeping 99.2% of the necessary connections between steps and 98.7% of the safety checks reachable.
The results were impressive. When using SkillZip, the robots solved tasks much better than when using older methods. On the data challenges, the robots improved their success rate by up to 12.2 points compared to the next best system. Even more importantly, the robots didn't just get faster; they got smarter. Because they weren't bogged down by reading irrelevant parts of the manuals, they made fewer mistakes and needed fewer attempts to finish their jobs. The system also showed it could grow: as the library of skills expanded from 200 to 100,000, SkillZip kept the robots from getting confused, proving that this "zipper" approach works even when the instruction book becomes a massive encyclopedia.
In short, SkillZip suggests that the future of robot intelligence isn't just about giving them more information, but about giving them the right information, compressed perfectly so they can find exactly what they need, when they need it, without ever losing the safety rules that keep them from crashing.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.