π± Claude Co-Work
Claude Co-Work is a further example of Claude moving from AI as chat assistant towards agentic behaviour (π± A shift from 'chat' assistants to action takers). It can be thought of as a βwork agentβ that can operate on your computer: opens files, browses the web, uses tools, runs code, and completes everyday tasks.
This video (February 9, 2026) describes Co-work through four mechanisms:
- Browser control (Chrome extension) β it can navigate pages, fill out forms, pull data from websites, and conduct βClaude researchβ by taking over browser actions.
- Connections / connectors β connecting Claude to external apps (he uses the term βMCPsβ and also shows a connectors screen). Examples mentioned: Google Sheets, Notion, Gmail, n8n, and others.
- Skills β add-on βinstruction manualsβ that Co-work reads before executing a task, intended to make outputs consistent and higher quality.
- File access β the ability to work in folders and manipulate real files (Word docs, spreadsheets, PDFs, etc).
Claude Co-Work can operate in two modes: either asking before acting, or acting without asking.
π± Claude Skills are an important component of Claude Co-Work.
The video mentioned above claims that if you accumulate enough skills, you can train Co-Work like a highly capable generalist worker with consistent results and delegated execution. I would advise caution before accepting this conclusion. A badly written Skill can easily lead to governance and quality risks, and the claimed consistency relies on the quality of the Skills and prompts.
Claude Co-Work moves Claude closer to being an agent (see π± Claude Skills vs Agents) but falls short of a full agentic framework due to a lack of full orchestration, schedules, etc.