Claude Cowork: the interesting part isn't the demo

Anthropic launched Cowork in January. The short version: it's the same agent that powers Claude Code, repackaged for non-developers and pointed at your local files. You give it a folder and a goal. It works.

Most of the writing about it has focused on the demos — organising messy downloads, drafting reports from a folder of sources, pulling expense data out of receipt screenshots. The demos are fine. They're not the point.

What it actually is

Cowork is a research preview that runs inside the Claude desktop app on macOS and Windows. It's available on paid plans — Pro, Max, Team, Enterprise. You scope it to a folder, describe the outcome, and it executes multi-step work: reading files, writing files, calling connectors, occasionally driving the browser. Deletions need explicit permission. You can stop it at any step.

Mechanically it's the agent architecture from Claude Code with the terminal stripped away and a sandboxed filesystem mounted underneath. Anthropic built it in roughly two weeks, reportedly using Claude Code itself to do most of the work.

What's new, and what isn't

The capability isn't new. Anyone with a terminal habit has been pointing Claude Code at non-coding folders for a year. What's new is the audience.

The question that matters: does an agent in a familiar desktop app, with no terminal, change which knowledge workers can actually use this? Our view is yes — meaningfully. The gap between "this is technically possible" and "the finance team will use it on a Tuesday" has been the bottleneck on agent adoption in the businesses we work with. Cowork narrows that gap by enough to matter.

Where we'd point it first

The work that pays inside an operating business tends to share a shape. High-volume, document-heavy, rules-bounded, currently swallowing real hours from people who'd rather be doing something else. The candidates worth piloting:

• Inbox triage on a shared mailbox — classify, draft replies, escalate the genuine exceptions
• File-system reconciliation across a deal room, a shared drive, and a CRM
• First-pass review of contracts or supplier docs against a checklist
• Pulling structured data out of unstructured sources — invoices, statements, forms
• Drafting board packs or client reports from a folder of source files

None of these are interesting. All of them have meaningful hours behind them. The boring tasks are where the early margin is.

What we'd be cautious about

Three things, in our view.

The agent runs locally and the data stays in front of it. That's the design and it's the right design — but it means scope matters. Don't point Cowork at a folder containing material you wouldn't want a contractor to read. The mitigation is operational, not technical: scope the folder, scope the connectors, scope the browser permissions.

Prompt injection remains real. If the agent reads a file or visits a page that contains adversarial instructions, the model can be steered. Anthropic has built defences and is candid about the limits. For sensitive workflows — anything touching customer data, financials, or external comms — keep a human in the approval loop, not just the review loop.

And the productionisation gap is still where most projects die. Cowork makes the first demo cheap. Turning the demo into a workflow your team uses every Wednesday is a different exercise, and it's the one most teams underestimate. Standard operating procedures, error handling, escalation paths, change management. The agent is the easy bit.

The honest take

Cowork is a meaningful piece of infrastructure for the kind of work we help businesses automate. It's not a strategy. It's a tool that makes certain strategies tractable that weren't tractable before.

If you've been waiting for the agent layer to be usable by non-developers before piloting anything: that wait is over. If you're hoping the tool itself will tell you which workflow to point it at: it won't.

That part is still the work.