PickleeAI와 개발, 오늘 볼 변화만
홈읽을거리아카이브
검색

Picklee

AI와 개발 현장에서 오늘 확인할 변화만 선별합니다.

© 2026 Picklee. All rights reserved.

RSSSitemap

읽을거리

2026년 7월 4일

Claude Science - Anthropic turns scientific workflows into auditable agent workbenches

Anthropic이 macOS, Linux, SSH/HPC 환경에서 동작하는 Claude Science beta를 공개했다. 이 워크벤치는 PubMed, Jupyter, R, cluster terminal 등으로 분산된 과학 연구 도구를 통합하고, 60개 이상 curated skills/connectors, reviewer agent, reproducible artifact history를…

본문 읽기원문 보기

발행일

2026년 7월 4일

업데이트

2026년 7월 4일

주제

AI
Claude
에이전트
원문 보기

배경 및 맥락

과학 연구에서 AI 활용은 논문 요약, 코드 보조, 데이터 분석 자동화로 시작했지만 실제 병목은 훨씬 넓다. 연구자는 PubMed, Jupyter, R, cluster terminal, 생물학 데이터베이스, 시각화 도구, manuscript editor를 오가며 작업하고, 결과 재현성과 citation 검증까지 관리해야 한다.

Claude Science는 이 흩어진 과학 워크플로를 agent workbench로 묶으려는 시도다. 같은 주간의 Claude 모델·정책 뉴스와 달리, 이 발표의 핵심은 모델 성능보다 domain-specific workflow와 auditable artifacts다.


핵심 내용

Anthropic은 2026년 6월 30일 Claude Science beta를 Pro, Max, Team, Enterprise 사용자를 대상으로 공개했다. 앱은 macOS와 Linux에서 동작하고, 원격 machine over SSH 또는 HPC login node에서도 사용할 수 있다.

기능적으로는 60개 이상 curated skills와 connectors를 제공하며 genomics, single-cell, proteomics, structural biology, cheminformatics 영역을 겨냥한다. Claude Science는 3D protein structures, genome browser tracks, chemical structures 같은 rich scientific artifacts를 native로 렌더링하고, figure 생성 시 code, environment, plain-language description, message history를 함께 남긴다. reviewer agent는 citation, calculation, figure-code mismatch를 점검한다.


경쟁 구도 / 비교

기존 Jupyter Notebook은 재현 가능한 분석의 기본 단위였지만, 데이터 검색, HPC job orchestration, 전문 데이터베이스 쿼리, manuscript refinement까지 자연어 agent가 연결해주지는 않는다. Claude Science는 notebook, cluster workflow, domain connectors, AI reviewer를 하나의 연구 세션으로 통합하려 한다.

Google DeepMind, OpenAI, NVIDIA BioNeMo, specialized biotech AI startups도 AI for Science를 겨냥하지만, Anthropic의 차별점은 연구자가 이미 가진 infrastructure와 proprietary tools를 connector와 skill로 붙이는 workflow surface에 있다.


의미

AI for Science의 실무 경쟁력은 답변 품질보다 재현성, 감사 가능성, 데이터 경계 관리에서 나온다. Claude Science가 로컬/HPC에 데이터와 compute를 두고 필요한 context만 모델에 보내는 구조를 강조한 것은 regulated research와 민감 데이터 환경에서 중요하다.

연구 리더는 도입 전 실험 설계, 데이터 접근권, compute budget, reviewer workflow, artifact retention policy를 함께 정해야 한다. agent가 생성한 figure나 분석 결과가 출판·임상·투자 의사결정에 쓰이려면, 모든 산출물의 lineage와 independent validation이 필수다.

이어 읽기

관련 읽을거리

전체 보기
2026년 6월 25일Claude Tag - Slack-native team agents move from private assistants to shared workspacesAnthropic은 Slack에서 @Claude를 태그해 팀 단위로 작업을 위임하는 Claude Tag beta를 공개했다. Claude Enterprise와 Team 고객 대상이며, channel-scoped memory, tool/data/codebase access, ambient updates, spend limits, activity logs를 제공한다.2026년 5월 26일Anthropic acquires Stainless — 에이전트 경쟁의 병목이 모델 지능에서 connector supply chain 통제로 이동Anthropic은 2026년 5월 18일 SDK·CLI·MCP server 생성 툴링 회사 Stainless를 인수한다고 발표했다. Anthropic은 Stainless가 자사 공식 SDK를 초기부터 생성해 왔으며, 수백 개 기업이 Stainless를 사용해 TypeScript·Python·Go·Java 등 다수 언어용 SDK와 MCP server를 만들고 있다고 설명했다.2026년 5월 19일Claude Managed Agents updates — self-hosted sandbox와 MCP tunnel이 enterprise agent 경계를 재정의Anthropic은 2026년 5월 19일 Claude Managed Agents에 self-hosted sandboxes와 MCP tunnels를 추가했다고 발표했다. self-hosted sandbox는 public beta, MCP tunnels는 research preview로 제공되며, agent loop는 Anthropic 인프라에 남기되 실제 tool execution과…2026년 5월 18일Anthropic acquires Stainless — 에이전트 경쟁의 병목이 모델 성능에서 SDK·MCP 연결성으로 이동Anthropic은 2026년 5월 18일 SDK 및 MCP server tooling 회사 Stainless를 인수한다고 발표했다. Stainless는 2022년 설립 이후 Anthropic의 공식 SDK 전부를 생성해 왔고, API spec으로부터 TypeScript, Python, Go, Java, Kotlin 등의 SDK, CLI, MCP server를 만드는 툴링을 제공해 왔다.