SkillOrchestra: Learning to Route Agents via Skill Transfer
Jiayu Wang, Yifei Ming, Zixuan Ke, Shafiq Joty, Aws Albarghouthi, Frederic Sala · 2026
Read SkillOrchestra: Learning to Route Agents via Skill Transfer by Jiayu Wang, Yifei Ming, Zixuan Ke, Shafiq Joty, Aws Albarghouthi, Frederic Sala (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-2602-19672.txt.
Description
Compound AI systems beat single models only when orchestration works, yet today's routers decide too coarsely or cost too much to adapt. Jiayu Wang and five co-authors present SkillOrchestra, a skill-aware orchestration framework that learns fine-grained skills from run experience instead of a routing policy end to end. It models each agent's competence and cost per skill, then matches skill demands to agents under an explicit performance-cost trade-off.
Across ten benchmarks, SkillOrchestra beats state-of-the-art RL orchestrators by up to 22.5 percent, while cutting learning cost 700-fold against Router-R1 and 300-fold against ToolOrchestra. Readers get a scalable, interpretable, sample-efficient alternative to data-hungry RL routing, with code available.
Catalog record
| CREATOR | Jiayu Wang, Yifei Ming, Zixuan Ke, Shafiq Joty, Aws Albarghouthi, Frederic Sala |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 120 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-19672