AI
SkillSeek: Revisiting Agent Skill Retrieval at Marketplace Scale
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 29/09/2026, 20:00:00 UTC
Briefing created: 01/10/2026, 08:53:46 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 01/10/2026, 04:18:24 UTC.
Briefings
Anthropic's Agent Skills package aggregates over 230,000 skills, creating a bottleneck where selection rather than authoring dominates. The literature currently outsources skill selection to agents via LLM-mediated retrieval loops, which incur high token costs. This briefing introduces SkillSeek, an open-source two-stage retriever using a standard IR recipe (BGE-base bi-encoder + small cross-encoder) via MCP. Evaluated on 89 tasks of SkillsBench, SkillSeek matches the performance of Liu et al.'s LLM loop at minimal extra cost, with plain bm25 matching the LLM loop on three settings and a cross-encoder covering the fourth. Total per-trial spend dropped from USD 51.30 to USD 27.54, positioning the IR recipe as a strong default for agent-skill retrieval.
Key points
- Anthropic's Agent Skills package aggregates over 230,000 skills, creating a bottleneck where selection rather than authoring dominates.
- SkillSeek uses a two-stage retriever with BGE-base bi-encoder and small cross-encoder via MCP to match LLM loop performance at minimal extra cost.
Context and limitations
Evaluation was conducted on 89 tasks of SkillsBench using OpenHands harness; results are limited to this specific benchmark and task set.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/10/2026, 08:53:46 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifaktdst3cm2nwy6dqx5xqc52dfngpj5u6jg5alinv4tfrsceayem