배경 및 맥락
AI coding agent와 업무 agent는 지금까지 개인 IDE, CLI, chat tab 중심으로 확산됐다. 이 구조는 개인 생산성에는 강하지만, 팀 전체가 같은 agent state를 보고 이어받거나, 특정 channel의 tacit knowledge를 축적하거나, 조직 차원의 권한과 비용을 통제하는 데 한계가 있다.
최근 cache의 Google Jules evals는 coding agent가 무엇을 알려야 하는지 평가하는 insight-policy 문제를 다뤘고, huggingface_hub release CI는 release automation에서 human review와 deterministic manifest가 중요하다는 점을 보여줬다. Claude Tag는 평가나 CI가 아니라 협업 표면 자체를 Slack channel로 옮기는 제품 방향이다.
핵심 내용
Anthropic은 2026년 6월 23일 Claude Tag를 공개했다. Claude Tag는 Slack channel 안에서 @Claude를 호출해 tasks를 위임하는 방식으로 동작하며, Claude가 접근 권한을 받은 tools, data, codebases를 사용해 단계별로 작업하고 Slack thread에 결과를 남긴다. beta는 Claude Enterprise와 Team 고객에게 제공된다.
핵심 기능은 multiplayer interaction, channel-scoped learning, ambient behavior, asynchronous work다. 같은 Slack channel 안의 사용자는 하나의 Claude context를 공유하며, Claude는 권한이 부여된 channel과 data source에서 관련 context를 기억한다. ambient mode가 켜져 있으면 관련 업데이트를 먼저 알려주거나 unresolved task를 follow-up할 수 있고, 장시간 작업을 스스로 schedule해 hours or days 단위로 진행할 수도 있다.
관리 기능도 중요하다. 관리자는 channel별 tools와 data access를 지정하고, 서로 다른 Claude identity의 memory를 scope로 분리할 수 있다. organization과 channel별 token spend limit, activity log, requester tracking도 제공된다. Anthropic은 기존 Claude in Slack app을 Claude Tag로 대체하고, 30일 이내 migration opt-in 경로를 제공한다고 설명했다.
경쟁 구도 / 비교
Claude Tag는 GitHub Copilot이나 Claude Code 같은 개인 coding assistant와 다르다. Slack channel의 shared context와 team-visible execution log를 중심으로 하므로, agent가 개인 도구에서 collaborative work actor로 이동한다. 이는 Salesforce-Fin acquisition처럼 customer service agent가 CRM suite에 편입되는 흐름과도 닮았지만, Claude Tag는 특정 업무 도메인이 아니라 horizontal collaboration layer를 노린다.
Arcade Series A 항목이 enterprise agent authorization layer를 강조했다면, Claude Tag는 authorization을 실제 product surface에 묶는다. channel별 memory isolation, tool access, spend control이 없으면 Slack-native agent는 빠르게 편해지는 만큼 데이터 경계와 비용 경계도 흐려질 수 있다.
의미
산업적으로 AI agent의 다음 경쟁 표면은 모델 성능만이 아니라 shared workspace integration이다. 업무는 개인 prompt보다 Slack thread, issue tracker, docs, dashboard, codebase 사이에서 일어나므로 agent가 그 맥락을 지속적으로 읽고 팀이 함께 볼 수 있는 방식으로 행동해야 한다.
실무적으로는 Claude Tag 도입을 단순 Slack bot rollout로 보면 위험하다. channel taxonomy, data sensitivity, tool permission, memory retention, token budget, audit responsibility를 먼저 정해야 한다. 특히 engineering, sales, finance처럼 권한 경계가 다른 조직은 Claude identity와 channel scope를 분리하고, agent output을 승인 없이 실행하지 않는 policy를 명확히 해야 한다.