skill-routing
daily-paper-2026-07-09-retriever-vs-decomposition-skill-routing
Retriever Bottleneck vs. Decomposition
TL;DR — In a real 1,898-skill bilingual agent harness, the ceiling-gap diagnostic
shows that neither decomposition nor a better retriever raises routing coverage: the
binding constraint is skill-corpus redundancy. Decomposition still earns its place on
execution ordering.
ThakiCloud AI Research · 2026-07-09 (v2, revised) · 📝 Tech blog (KO)
Problem
Operators of large agent harnesses that route natural-language requests to… See the full description on the dataset page: https://huggingface.co/datasets/thaki-AI/daily-paper-2026-07-09-retriever-vs-decomposition-skill-routing.daily-paper-2026-08-31-multi-skill-gap-sequence-routing
The Multi-Skill Gap: Measuring the Cost-Quality Frontier of Order-Sensitive Skill-Sequence Routing in a 2,200-Skill Agent Harness
TL;DR — First order-sensitive chain benchmark for skill routing: 48 composite tasks (12 real workflow templates x 4 variants) over a 2,275-skill bilingual production registry, measured across free BM25 top-k retrieval, self-hosted Qwen3-1.7B/14B composers, and a frontier cost anchor. Free top-20 retrieval reaches 75.7% set completeness but 0.0% exact… See the full description on the dataset page: https://huggingface.co/datasets/thaki-AI/daily-paper-2026-08-31-multi-skill-gap-sequence-routing.daily-paper-2026-09-04-redundancy-effect-skill-registry-routing
The Redundancy Effect: Decomposing Skill-Registry Size from Duplicate Mass in Router Accuracy of a 2,200-Skill Agent Harness
TL;DR — In a ~2,200-skill agent harness, much of the strict top-1 routing-accuracy loss seen as registries grow is a scoring artifact of near-duplicate skills rather than router degradation: a closed-form symmetry result shows that at fixed quality strict top-1 decays as 1/(m+1) in the number of redundant copies while Recall@k (k >= m+1) is unaffected -… See the full description on the dataset page: https://huggingface.co/datasets/thaki-AI/daily-paper-2026-09-04-redundancy-effect-skill-registry-routing.daily-paper-2026-09-13-selective-skill-routing-veto-frontier
The Veto Frontier: Measuring Skill-Router Calibration and the Asymmetric Cost of Abstaining vs. Auto-Routing in a 2,000-Skill Agent Harness
TL;DR — A production skill router that must pick one of ~2,000 skills every turn holds only one confidence signal: the RRF margin between the top-1 and the runner-up. This paper builds the economics of that margin under the asymmetric loss where a wrong auto-route wastes the entire downstream task while escalating to an LLM judge costs one… See the full description on the dataset page: https://huggingface.co/datasets/thaki-AI/daily-paper-2026-09-13-selective-skill-routing-veto-frontier.cortex-skill-routing