Skill++ rewrites new skill metadata so agent routers stop displacing older skills
Researchers from UIC, McGill and other institutions propose Skill++, which rewrites a new skill's metadata so agent routers keep choosing the right existing skills.
Researchers from the University of Illinois Chicago, McGill University and several other institutions have proposed Skill++. The method rewrites the name and description of a new agent skill before it joins a library, so that the router keeps selecting older skills it used to get right. The authors, including Hanrong Zhang and Philip S. Yu, frame the problem as lifelong skill routing. In their account, the router is the component that decides which skill an agent should use for a given request. Skill++ does not retrain the router.
The mechanism depends on how agents load skills. Under progressive disclosure, an LLM router sees only each skill's name and description, and loads the full body only for the skills it selects. Because the router relies on that short metadata, a new skill whose description resembles an earlier one can take over queries the earlier skill used to handle. This happens even though neither the old skill nor the router has changed. The arXiv abstract says this interference is worse for public skills, whose descriptions may not match their bodies.
The authors argue that existing fixes fall short. Continually retraining the router risks catastrophic forgetting, the loss of earlier behavior when a model learns new material. According to the abstract, rewriting methods either ignore similar skills or require feedback on each skill. Skill++ instead retrieves the most similar earlier skills and writes a new description. That description states what the skill's body supports, how it differs from its closest predecessor, and what related work it excludes. The method needs no router training and no labeled data, and it changes only the metadata, not the skill body.
In the growing-library setting, the router selects every required skill for 76.3% of queries with Skill++, against 36.4% with the original metadata. The DAIR.AI Academy write-up of the paper says the rewriting nearly eliminates substitution errors and halves crowding errors. The authors report an average gain of 23.1 points in final-session routing accuracy across three benchmarks, SRA-Bench, SkillRet and SkillRouter, and five LLM routers. They say the method beats all compared approaches. The gains also carried over to the Codex CLI agent harness.
These figures come from the authors' own benchmarks and comparisons. The 76.3% number describes one growing-library setting, not general routing accuracy. The abstract does not name the five LLM routers, so readers cannot tell from that page which routers the average covers. The method starts by retrieving the most similar earlier skills to compare against, and the abstract does not analyze how well that retrieval works.
DAIR.AI's commentary draws a practical lesson from the paper. Maintainers of a skill library can apply the core idea by writing down how each new skill differs from the closest one already installed. That advice is a manual version of what Skill++ automates, and it targets the same gap between skill bodies and the metadata the router reads.
What changed
Skill++ rewrites a new skill's name and description before it joins a library so the router keeps selecting older skills, without retraining the router.
Type: Research
Why it matters
Companies maintaining growing skill libraries may see routing errors rise as libraries expand, and poorly written skill descriptions can make interference worse.
What you can do
When adding a skill to a library, describe how it differs from the closest installed skill.
Original source
Where we saw it
DAIR.AI (Elvis Saravia)
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works