Vibe Working: What It Actually Means When AI Does Your Job

Anthropic’s Scott White, Head of Product for Enterprise, used a specific phrase in late January 2026 that stuck: “vibe working.” Not vibe coding — that’s Andrej Karpathy’s term for developers who direct AI to write code without touching it themselves. Vibe working is the same idea extended to knowledge work broadly. Lawyers directing AI through contract review. Financial analysts having AI run through earnings calls. HR teams letting an agent draft policies, flag compliance gaps, and format the output for leadership. The person in charge isn’t doing the task — they’re steering something that does it for them. That shift is happening faster than most organizations have noticed, and Anthropic is betting a significant portion of their product roadmap on it becoming the default mode of professional work.

What Claude Cowork Actually Is

Claude Cowork launched in research preview at the end of January 2026. It’s a desktop app — macOS first — that runs Claude inside an isolated virtual machine on your local computer. That matters for a specific reason: it has full access to your local files and MCP (Model Context Protocol) integrations without routing everything through a cloud API call every time. The agent can open a file, work through it, reference another file, and iterate — all in a persistent thread that doesn’t reset when you close the chat window.

That persistence is the thing. Most AI work today is stateless and episodic. You open a chat, do something, close it. Cowork treats Claude more like a colleague with ongoing context about your work. Pro and Max plan users get a persistent agent thread accessible from both mobile and desktop. If you start a research task on your laptop in the morning and want to check progress from your phone at lunch, the thread is still there, still working.

Cowork ships with domain-specific plugins for legal, financial analysis, HR, engineering, and operations work. These aren’t cosmetic rebranding — they’re structured access patterns for how Claude approaches different professional contexts. A legal plugin understands you’re working with privileged documents. A financial analysis plugin knows you’re looking at structured data that needs to be cited carefully.

One detail that says a lot about the pace Anthropic is moving at: Cowork itself was built using Claude Code in ten days. Anthropic engineers are now using Claude for roughly 60% of their own work and shipping 60 to 100 internal releases per day. That’s not a marketing claim — it’s the logical output of a team eating their own product compulsively.

The Capability Stack Underneath the Interface

Cowork runs on Claude Opus 4.6 and Sonnet 4.6, and understanding what those models actually do changes how seriously you take the “vibe working” framing.

Claude Opus 4.6 carries a 1 million token context window — available by default for Max, Team, and Enterprise plan users. To put that concretely: 1 million tokens is roughly 750,000 words, or the equivalent of loading an entire multi-year contract history, a full codebase, or a year of financial filings into a single working context. Anthropic’s own Frontier Red Team used Opus 4.6 to find more than 500 vulnerabilities in production open-source code. That’s a security team using the model as an active research instrument, not a search assistant.

Claude Sonnet 4.6 launched on February 17, 2026, at the same price as 4.5 with improved performance, better agentic search, and more efficient token usage. The 1 million token context window is in beta for Sonnet. For most professional knowledge work — documents, research, drafting, analysis — Sonnet 4.6 is the practical everyday model. Opus 4.6 is for the heavy lifts: the tasks where context breadth is the constraint.

Both earlier models — Opus 4 and Opus 4.1 — have been removed from the model selector. Anthropic is consolidating rather than accumulating. That’s a signal about where their confidence is placed.

Below the model layer, Claude Code has been shipping daily releases with features that read like a checklist for turning AI into an actual workflow system rather than a chat toy:

• Agent Skills / Skills API: Organized folders with SKILL.md files that teach Claude a repeatable task — generating a PPTX deck from a research doc, formatting XLSX output, building a DOCX summary from raw notes, processing PDFs into structured data
• Pre-built skills for PPTX, XLSX, DOCX, and PDF out of the box
• Voice mode for hands-free direction
• Worktrees for parallel task management
• –channels permission relay (research preview) for multi-agent coordination
• –bare flag for scripted automation pipelines
• MCP tool improvements across the board

Claude Code is now included in every Team plan standard seat. Claude also exists as a Chrome extension that reads console errors, DOM structure, and network requests in real time. There are add-ins for Excel — including pivot table editing, conditional formatting, and Opus 4.6 access — and a PowerPoint add-in as well. The surface area of where Claude can operate is expanding faster than most people tracking it from the outside realize.

Who Actually Benefits, and How

The vibe working framing can sound like an HR headline — “AI does your job for you!” — when the reality is more specific and more interesting. The leverage is asymmetric based on what kind of knowledge work you do and what proportion of it is high-judgment versus high-volume.

Here’s a practical breakdown:

The pattern across all of these: AI absorbs the volume work that used to fill hours, and the human operates at the level of direction, judgment, and accountability. That’s not the AI “doing your job.” It’s more like having a capable, tireless analyst who never gets bored of reading documents but needs your judgment to know what matters.