Claude Cowork Goes Mobile: Anthropic's Quiet Bet That Agents Belong Everywhere You Work

The most interesting thing about Anthropic’s move to bring Claude Cowork to mobile and web isn’t that you can now poke an agent from your phone. It’s the admission buried in the announcement: more than 90 percent of Cowork usage, according to Anthropic, has nothing to do with code. The company built an agentic harness that looked like a developer tool, and finance leads, ops managers, and content teams grabbed it instead. Anthropic built access where those users already work.

Why Mobile and Web Access Changes the Agent Loop

Anthropuc is expanding Claude Cowork beyond its desktop-only beta, starting the rollout with Max subscribers and widening from there over the coming weeks. Tasks kicked off at a desk can now be checked from a phone and reopened in a browser anywhere else, with Claude continuing to run in the background even when the laptop is closed or the phone is off.

Agents that only live on desktop are agents you have to babysit at a workstation, which defeats most of the point of asynchronous work. Making the runtime portable turns Cowork into something closer to a Slack thread you can walk away from. If you’re an operations lead who kicked off a quarterly spend reconciliation before a meeting, you can now get pinged on your phone the moment Claude hits a decision it can’t make alone, approve or redirect, and let it keep going. The editorial read: this is Anthropic conceding that the desktop app was a distribution bottleneck, not a design choice, and the sooner they fix it the sooner Cowork stops looking like a power-user toy.

The 90 Percent Non-Coding Signal Is the Real Story

Buried in the announcement is the fact that more than 90 percent of Cowork usage isn’t software work, per Anthropic, with business operations and content creation making up roughly half of all usage between them. That’s not a rounding error — that’s a product identity crisis with a happy ending.

For years the conversation around agentic tooling has been dominated by coding agents — Cursor, Copilot, Codex, Devin. Anthropic just published field data suggesting the general knowledge-work agent market is bigger than the coding one, at least inside their own user base. If you’re building internal tooling, this reframes the question. Instead of asking “can we get an AI to write our code,” ask whether you can hand off recurring artifacts — variance memos, client decks built from call transcripts, weekly ops summaries — to something that runs in the background. That’s the boundary where teams weighing AI agents versus more deterministic AI automation actually have to make a choice, and Cowork’s usage mix pushes toward agents for open-ended work. The prediction: within a year, at least one major analyst report will formally split “coding agents” and “knowledge-work agents” into separate categories, because pretending they’re the same market is starting to hide too much signal.

Human-in-the-Loop Was the Missing Mobile Feature

When Claude hits a point where only a person can make the call, the agent now asks via smartphone, and Anthropic says nothing gets sent without user review and approval. That design decision is what makes mobile viable at all.

Most autonomous-agent failures in production aren’t runaway hallucinations — they’re the agent stalling because it needs a decision no one is around to make. Pushing approval requests to the device people actually carry closes that gap. Imagine you’re running a small consultancy and you’ve asked Claude to draft a client proposal from a discovery call transcript. It gets 80 percent through, hits a pricing question, pings your phone at 4:47 PM, you tap “go with tier 2,” and the deck is done before you’re home. That’s when agents stop being a science project. Teams commissioning custom AI agent builds should prioritize this pattern — mobile approval is going to be table stakes, not a differentiator.

The Chat and Cowork Merger Is Coming

On web and desktop, Chat and Cowork will now share a single home screen, projects and artifacts will be available across platforms, and Anthropic is extending its doubled Cowork usage limits through August 5. The direction is clear: apart from local file access, which the web version lacks, the wall between Chat and Cowork is coming down.

Anthropuc isn’t the only one doing this. OpenAI has publicly floated merging ChatGPT, Codex, and the Atlas browser into a single desktop superapp, and Mistral already pulled the trigger, replacing Le Chat with Vibes as its agentic coding environment. What changes practically? Users will stop having to decide up front whether their question is a “chat” or a “task.” You’ll type something into a single box, and the product will decide whether to answer inline or spin up a background agent. That’s a real usability gain, but it also collapses pricing tiers and makes it harder for enterprises to budget agent usage. For decision-makers weighing custom AI builds against off-the-shelf SaaS AI, the merger trend runs in both directions: the SaaS options get more capable but harder to reason about cost-wise. The prediction: by mid-2027, “Cowork” as a distinct brand will be gone, folded into Claude the way Codex is likely to be folded into ChatGPT.

FAQ

Q: What is Claude Cowork? A: Claude Cowork is Anthropic’s agentic harness for general knowledge work — a Claude mode that can plan, execute, and iterate on multi-step tasks in the background rather than just answering in a single turn. Anthropic reports that more than 90 percent of its usage is non-coding work, dominated by business operations and content creation.

Q: Do I still need the desktop app? A: Yes, if you want local file access, connected folders, browser control through Claude in Chrome, or Computer Use where Claude clicks and types on your screen directly. The mobile and web versions cover everything else, and users who couldn’t install the desktop app can now try Cowork through a browser for the first time.

Q: How is this different from ChatGPT’s agent mode? A: Functionally they’re converging. OpenAI plans to merge ChatGPT, Codex, and Atlas into a single desktop superapp, and Mistral already replaced Le Chat with Vibes. Cowork’s mobile expansion and shared home screen with Chat put Anthropic on the same path — a single conversational surface that transparently decides when to run a background agent.

Key Takeaways

• Teams building internal AI workflows should stop assuming “agent” means “coding agent” — Anthropic’s own data says the general knowledge-work use case is larger.
• Mobile approval flows are becoming the difference between agents that ship value and agents that stall waiting for input; any custom agent build should design around this pattern from day one.
• The Chat-versus-Cowork distinction is going to disappear across every major vendor within 18 months; product roadmaps that depend on it are already stale.
• Desktop is not dead for agents — local file access, Computer Use, and browser control remain desktop-exclusive, so hybrid deployments will be the norm for at least the next two years.
• Watch for enterprise pricing to get messier as background agent usage becomes indistinguishable from chat usage; budget planners should push vendors for usage transparency now, before the merger happens.