Retrieval-Augmented Skill Optimization via Cross-Harness Adaptation
Jaewon Chu, Ji Soo Lee, Jihwan Park, Dohwan Ko, Jeehye Na, Seunghun Lee, Taehoon Lee, Minseo Yoon and 3 others
Korea University · KAIST · Meta AI
Published September 29, 2026 · on the feed October 2, 2026 · ▲ 60 · arXiv 2609.38024
The abstract
An agent skill is a reusable, actionable natural-language artifact that guides an agent to perform a task effectively under a given harness. Recent studies have explored the optimization of agent skills, contributing to a growing collection of publicly available skills spanning diverse tasks, domains, and harnesses. Despite millions of publicly shared skills, existing skill optimization methods largely overlook this accumulated knowledge, instead relying solely on expensive agent rollouts to iteratively refine skills for a target task. To address this, we propose Retrieval-Augmented Skill Optimization (RASO), a framework that leverages an external skill corpus as prior knowledge throughout skill optimization. RASO retrieves relevant knowledge from existing skills and adapts it to the target task and harness via Cross-Harness Adaptation, accounting for mismatches in both domain and harness. RASO comprises two complementary stages: Retrieval-Augmented Skill Initialization (RASI) constructs a knowledge-grounded initial skill without requiring agent rollouts, while Retrieval-Augmented Skill Update (RASU) iteratively refines the skill by retrieving external knowledge guided by execution feedback. Across four agent benchmarks and two models, extensive experiments show that RASO consistently outperforms baselines without retrieval-augmented skill initialization and updating.
The Brief
2:25A short podcast summary of the paper. Written and voiced by AI.
Transcript
Retrieval-Augmented Skill Optimization via Cross-Harness Adaptation. By Jaewon Chu and ten others, from Korea University, KAIST, and one other institution. Published on the twenty-ninth of September. No code repository is linked to it.
Start with the object being optimized. An agent skill is a reusable, actionable artifact written in natural language, and its job is to guide an agent to perform a task effectively under a given harness. The harness matters: the same task under a different harness may want a different skill.
Recent studies have explored how to optimize these skills, and alongside that work a large public collection has grown up, spanning diverse tasks, domains, and harnesses. There are millions of shared skills. Existing optimization methods, though, largely overlook all of it. They rely solely on agent rollouts, running the agent and refining the skill for the target task over and over, and those rollouts are expensive. The accumulated knowledge sits unused while each new skill is worked out the costly way.
Retrieval-Augmented Skill Optimization, RASO, puts that collection to work. It treats an external skill corpus as prior knowledge throughout the whole optimization, not only at one step. It retrieves relevant knowledge from existing skills, then adapts it to the target task and harness through what the authors call Cross-Harness Adaptation. Retrieved skills will rarely match the target exactly, and that adaptation accounts for mismatches of two kinds: in the domain, and in the harness.
RASO runs in two complementary stages. The first is Retrieval-Augmented Skill Initialization, or RASI. It constructs a knowledge-grounded initial skill, and it does so without requiring any agent rollouts, so the starting point costs nothing in agent execution.
The second stage is Retrieval-Augmented Skill Update, or RASU. It refines the skill iteratively, and at each round it retrieves further external knowledge, with the retrieval guided by execution feedback. What happens when the agent runs shapes what RASO looks up next.
Experiments span four agent benchmarks and two models. Across them, RASO consistently outperforms baselines that lack retrieval-augmented skill initialization and updating.
Where earlier optimizers leaned on rollouts alone and paid for every improvement, RASO starts from what others have already written down. The corpus supplies a grounded first draft before a single rollout, and execution feedback decides which part of that corpus to consult as refinement goes on.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Millions of written skills exist, and each new one is still worked out the expensive way. That is the wall.
You run a language-model agent inside a harness that sets its tools, file access and scoring. A short written guide on how to do the task, a skill, makes the agent far better.
The road ahead
- 1 · The problem ~7 min A short written guide, a skill, makes an agent much better, but improving that guide by repeated agent runs is slow and expensive.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Retrieve know-how from public skills, rewrite it for your own task and harness, and use it both to write the first skill and to fix later failures.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Search public skill sections, rewrite them for the target harness, build a starting skill, then loop: run, find failures, retrieve and adapt again.
- 6 · The experiment ~6 min You’ll guess: Across four agent benchmarks and two models, how did RASO compare with baselines lacking retrieval-augmented skill initialization and updating?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.