이번 주 AI 트렌드 — 2026-10-18 ~ 2026-10-24
코딩 에이전트에 얹는 스킬과 하네스가 이번 주 관심의 중심이고, 에이전트에 로컬 맥락과 도구를 공급하는 저장소가 함께 올라왔다.
뜨는 카테고리 · 식는 카테고리
| 카테고리 | 이번 주 비중 | 앞 4주 평균 | 변화 |
|---|---|---|---|
| 에이전트 스킬·프롬프트 | 26.7% | 23.9% | ▲ 2.8%p |
| 추론·서빙 | 3.7% | 1.5% | ▲ 2.3%p |
| 평가·관측(LLMOps) | 4.7% | 3.8% | ▲ 0.9%p |
| 기타 AI | 1.9% | 4.2% | ▼ 2.3%p |
| 브라우저·컴퓨터 사용 자동화 | 1.5% | 2.8% | ▼ 1.2%p |
| 코딩 에이전트·IDE | 6.4% | 7.5% | ▼ 1.1%p |
범위: 판정된 저장소(랭킹 후보 쪽)의 증가만 집계했다 — 이번 주 커버리지 48%.
새로 들어온 저장소
- Gathin23/cleanyourtv — Debloat any Android or Google TV over Wi-Fi and get an Apple TV-style home. · 에이전트 스킬·프롬프트
- deepseek-ai/DeepInfer — DeepInfer: High-performance, large-scale inference framework for DeepSeek models · 추론·서빙
- google-research/harness2 — google-research/harness2는 모델 가중치는 그대로 두고 프롬프트·도구·메모리·스킬 같은 에이전트 harness를 작업마다 재귀적으로 고치는 Google Cloud AI Research·Northwestern University의 연구용 Python 프레임워크다. · 에이전트 프레임워크
- YuJunZhiXue/dsh-purge — DeepSeek Harness 破甲:让所有模型都能破甲,不同模型可换不同提示词;默认提示词面向国模「小码酱」。Jailbreak for every model — swap prompts per model. 求 Star 收藏 ⭐ 如要二开,请查看署名要求部分 · MCP 서버·도구
- tangency-labs/solidcraft — SolidCraft: an open-source parametric 3D CAD application built in pure Rust. Desktop, web, and automation through MCP. · MCP 서버·도구
- gamedev-skills/awesome-gamedev-agent-skills — 74 game-dev skills for AI coding agents — Godot, Unity, Unreal, Phaser, PixiJS, three.js, Bevy, pygame, LÖVE, Roblox. Portable SKILL.md Agent Skills (the format Anthropic launched as Claude Skills), with a router that loads the right skill for your engine and task. Runs in Claude Code, Cursor, Kiro, Codex, Copilot, Gemini CLI and more. · 에이전트 스킬·프롬프트
- Folkbench/folkbench — AI API 中转站评测排行与切换工具:实测可用率、延迟与价格 · Open-source AI API relay benchmark, rankings & switcher for Claude Code / Codex · 평가·관측(LLMOps)
- lemomo-ai/lemo-opuscar — A toolbox of film styles made in code for Claude Code: 59 ready to use, and your own. Each style is a style prompt plus a demo film made entirely in code by Claude Opus 5.5, with no video model. By Lemomo. · 一套用代码做片的风格工具箱:59 种现成风格,也能做出你自己的 · 에이전트 스킬·프롬프트
- PC2005-cloud/dsh-pet — DSH 桌面宠物:一行命令装好即用的透明动画小桌宠,支持多开、大小位置随心配置;还内置 DIY 素材链,能用 AI 视频自造专属宠物 · MCP 서버·도구
- Vincentwei1021/mg-styles-15 — 15 motion design styles, each made by Claude Opus 5.5 writing code from one prompt: watch the films, copy the prompts, read the full source. · 에이전트 스킬·프롬프트
- Derpyu520/qq-bridge — Bridge between QQ (SnowLuma OneBot v11) and DeepSeek Harness agents: social simulation, safe MCP tools, slang learning and more. · AI 앱·서비스
- ethanplusai/harvey — Autonomous AI sales agent powered by Claude Code or Codex. Finds prospects, writes cold emails, sends campaigns, handles replies, books meetings. No API keys needed — runs on your Claude or Codex subscriptions. · AI 앱·서비스
- MerZlin/dsh-pet-indesktop — 基于项目PC2005-cloud/dsh-pet.git,将桌宠移植到Windows、Linux和MacOS上,现在可以随时看到蓝色大肥鱼了:),本项目还额外实现了其他交互功能,欢迎体验 · AI 앱·서비스
- amontlabs/lcu — Codex computer use, decoupled from the app, for usage inside any harness. · 브라우저·컴퓨터 사용 자동화
왜 지금 — 주목할 저장소
- google-research/harness2 — 모델 가중치는 그대로 두고 프롬프트·도구·메모리·스킬 같은 에이전트 하네스를 작업마다 재귀적으로 고치는 연구용 프레임워크다. arXiv 논문 공개 뒤 기사·HN 글·실행 영상이 같은 날 몰리며 급등했다. (확인)
- cloudflare/security-audit-skill — 코딩 에이전트를 6단계 보안 감사자로 돌리고, 각 단계를 격리된 에이전트가 맡아 검증된 기계 판독형 findings를 남기는 스킬이다. 9월 공개 뒤 해설이 이어졌고 셋업 가이드가 나온 뒤 다시 급등했지만, 급등 당일의 직접 계기는 확인되지 않는다. (확인)
- Graphify-Labs/graphify — 코드·문서·SQL 스키마를 tree-sitter AST로 로컬 파싱해 지식 그래프로 만들고 여러 코딩 에이전트에서 스킬로 쓰게 하는 도구다. 커뮤니티 PR을 묶은 릴리스가 거의 매일 나오고 실전 가이드가 공유되면서 관심이 다시 몰렸다. (확인)
시사점
- 에이전트 스킬·프롬프트 카테고리의 비중이 더 커졌다. 개별 프롬프트 모음을 넘어 보안 감사, 게임 개발, 엔지니어링 습관처럼 특정 작업 절차를 SKILL.md나 CLAUDE.md로 패키징해 여러 코딩 에이전트에서 재사용하는 방식이 주류로 굳어지고 있다. cloudflare/security-audit-skill · mattpocock/skills · gamedev-skills/awesome-gamedev-agent-skills · obra/superpowers · affaan-m/ECC · multica-ai/andrej-karpathy-skills (추정)
- 모델을 바꾸기보다 에이전트를 둘러싼 하네스를 다듬는 쪽으로 관심이 옮겨 가고 있다. 하네스를 자동 개선하는 연구 구현, 플러그인 중심 하네스, 컴퓨터 사용 기능을 하네스에 붙이려는 시도가 같은 주에 나란히 보인다. google-research/harness2 · deepseek-ai/deepseek-harness · affaan-m/ECC · amontlabs/lcu (추정)
- 벡터 스토어나 클라우드 API 없이 로컬에서 에이전트에 맥락과 기능을 주는 도구가 눈에 띈다. AST 그래프로 코드 구조를 주거나 로컬 검색 엔진을 MCP 서버로 붙이거나 음성 엔진을 로컬 API·MCP로 묶는 식이라, 사내 코드나 민감한 데이터를 다루는 팀이 시험해 볼 만하다. Graphify-Labs/graphify · xerj-org/xerj · debpalash/VoiceStudio (추정)
- DeepSeek Harness(dsh)를 겨냥한 플러그인·데스크톱 펫·메신저 브리지가 한꺼번에 들어왔다. 다만 카드 요약이 없어 설명만으로 판단해야 하고, 탈옥 프롬프트나 키 없이 무료 무제한 모델을 내세우는 항목도 섞여 있어 설치 전에 출처와 동작을 직접 확인하는 것이 안전하다. deepseek-ai/deepseek-harness · YuJunZhiXue/dsh-purge · PC2005-cloud/dsh-pet · Derpyu520/qq-bridge · MerZlin/dsh-pet-indesktop · Ebony-Vinyl/dsh-our-free-model (추정)
- 코딩 에이전트·IDE 비중은 줄었지만, 여러 에이전트를 격리된 worktree에서 병렬로 돌리고 추적하는 오케스트레이션 도구는 꾸준히 주목받는다. 추론·서빙과 평가·관측 비중이 오른 것도 함께 보면, 관심이 에이전트 자체에서 그 주변의 운영·인프라로 넓어지는 흐름으로 읽힌다. stablyai/orca · anomalyco/opencode · deepseek-ai/DeepInfer · Folkbench/folkbench (추정)
이번 주 범위
AI 탭 48개 중 카드가 있는 저장소 5개 · 처음 들어온 저장소 14개. 숫자는 코드가 계산했고, 글은 카드 요약과 저장소 설명만 근거로 썼다(새 검색 없음).
기준일 2026-10-24 · 인용 게이트 검사 33 · 제외 0