EvoSkill: Automated Skill Discovery for Multi-Agent Systems
Salaheddin Alzubi, [2pt] Noah Provenzano, Jaydon Bingham, Weiyuan Chen, Tu Vu, [2pt] Sentient, Virginia Tech · 2026
Read EvoSkill: Automated Skill Discovery for Multi-Agent Systems by Salaheddin Alzubi, [2pt] Noah Provenzano, Jaydon Bingham, Weiyuan Chen, Tu Vu, [2pt] Sentient, Virginia Tech (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-02766.txt.
Description
Salaheddin Alzubi, Noah Provenzano, Jaydon Bingham, and colleagues let coding agents grow their own reusable skills. Their 2026 framework, EvoSkill, starts from a real gap: flexible agents still lack deep know-how for specialized tasks. Hand-written skill packs help, but they do not scale.
EvoSkill studies an agent's failures, proposes a new skill or an edit to an old one, and writes it as a structured folder. Only changes that improve held-out test scores survive, judged against a frontier of best programs, while the base model stays frozen. Gains appeared on two question-answering benchmarks over treasury and search tasks. Skills learned on one task even transferred to a fresh browsing benchmark, lifting scores with no changes.
Catalog record
| CREATOR | Salaheddin Alzubi, [2pt] Noah Provenzano, Jaydon Bingham, Weiyuan Chen, Tu Vu, [2pt] Sentient, Virginia Tech |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 57854 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.02766