Anchor watch

“Anchor watch” is the least glamorous duty on a ship: one crewman stays on deck through the small hours while everyone else sleeps, checks that the anchor is holding and the ship hasn’t drifted, and wakes the captain only if something actually needs a decision. It is a strange thing to name a piece of enterprise software after, and it is exactly what Anthropic built. The interesting part isn’t that an AI agent can now work while you sleep — plenty of agents already claim that. It’s what Anthropic’s own numbers say those agents are actually being asked to do once they can.

01 What shipped

Claude Cowork launched in January as a desktop app, aimed squarely at people who don’t write code for a living — a Claude that could reach into your files, folders, and apps and run a multistep chore without being walked through it turn by turn. The catch, from day one, was that it needed your laptop. Close the lid mid-task and the task stopped with it.

On July 7, that constraint went away. Cowork sessions now default to running in the cloud: they follow you from a laptop to the Claude apps on the web, iOS, and Android, keep running when every device you own is powered off, and hand you a push notification the moment they need a decision only you can make. Scheduled tasks — the “draft this every Monday” kind — no longer need a machine to stay on to fire.

Access is deliberately narrow for now. Beta is limited to Anthropic’s Max subscribers, the $100-a-month tier, with usage limits doubled through August 5 before the company says it will widen access to more plans. Existing scheduled tasks don’t migrate automatically — they keep running the old, device-tethered way, while every new one created after July 7 defaults to the cloud.

02Who’s actually aboard

Anthropic published its own usage breakdown alongside the rollout: 1.2 million anonymized Cowork sessions, sampled May 11–31, drawn from more than 600,000 organizations. It’s the first real look at what people do with an agent once it’s allowed to run unsupervised, and the shape of it undercuts the “AI agents are for coding” assumption most of this industry still writes from.

Broken out by what the sessions were actually for:

The bucket Anthropic calls “the rest” is itself telling: research and intelligence work (6.4%), data analysis and business intelligence (5.8%), sales and revenue operations (4%), and legal and compliance (1.3%) — none of it individually large, all of it the kind of unglamorous, document-heavy work that fills most people’s actual calendars. A lawyer using Cowork to format and file paperwork, a hiring manager using it to schedule interviews and synthesize feedback, a team lead using it to turn notes into a slide deck: that’s the modal session, not an agent refactoring a codebase overnight.

Anthropic has been fairly direct about why this matters to the company’s own strategy. Claude Code proved the agent pattern works for a population of maybe thirty million developers worldwide. Cowork is the same pattern aimed at the much larger population of professionals whose job is creating, organizing, and communicating information rather than writing it in a programming language — call it the other four hundred million knowledge workers the coding-agent wars were never actually fighting over.

03 The watch bill

A watch bill is the roster that assigns each sailor a shift and a duty — the schedule that decides who’s allowed to act, and when someone has to wake the captain instead. Cowork’s equivalent isn’t a document, it’s a set of hard gates baked into what the agent is and isn’t allowed to do without you in the loop:

That line is doing a lot of work, and it’s worth taking seriously rather than treating it as marketing copy. The actual safety lever here isn’t a capability limit — Cowork can read your files, draft your correspondence, and reorganize your records without asking first. The lever is that the two irreversible actions, sending something outward and deleting something permanently, are the two things it’s structurally not allowed to do alone.

04 My read

The cloud move is the unsurprising half of this story — every serious agent product ends up here eventually, because “keep your laptop open” was never a real constraint, just a temporary one born of shipping the desktop version first. The usage data is the half worth sitting with. I’ve spent a fair amount of this year writing about agents from the coding-tool side of the fence — Claude Code, orchestration, the terminal — because that’s the side I work on. Anthropic’s own numbers are a useful corrective: the tool getting the cloud treatment first, ahead of anything coding-specific, is the one where 91% of the traffic was never about code at all. If you’re building for “AI agents” as a category and modeling the market on developer adoption curves, you’re sizing the smaller half of the room.

The approval gate is the part I’d watch most carefully, and not because I think it’s poorly designed — it’s a sensible, narrow chokepoint on exactly the two actions that can’t be undone. The risk is the oldest one in software UX: every “are you sure?” dialog in computing history degrades the same way, from a real decision point into a reflex tap once it fires often enough. A mobile push at 11pm asking you to approve a drafted email is a genuine safety mechanism the first dozen times. Whether it’s still one the three-hundredth time, once Cowork is running a dozen quiet sessions a day across an organization, is an empirical question nobody has data on yet — including Anthropic.