Anthropic expanded its agent product Claude Cowork to web and mobile (iOS and Android) on July 7, 2026. Cowork is an agent that works across connected tools such as files, calendar, email, messaging apps, and the web to carry a task through to completion, and what previously ran only on desktop now reaches the browser and the phone. Work sessions follow the user across devices, run in the background even when no device is active, and can be scheduled to fire on their own, such as a 6 a.m. client prep run. ASAP summarizes it from Anthropic's official announcement and primary reporting by TechCrunch and SiliconANGLE.

From desktop-only to three screens

What Anthropic loosened here is Cowork's reach, not a performance number. Cowork previously ran only in the desktop app, and this expansion lets the same product run on claude.ai in the browser and in the iOS and Android mobile apps. The point is not simply more screens but sessions that carry across device boundaries. A task started on a laptop can be picked up on a phone while on the move. Anthropic quoted Armmand Hosseini, Customer Success at Ramp: "I built a dashboard to track my clients while traveling. I started on my laptop and picked the session up on my phone while waiting for my bag to come out." Web and desktop now share one unified home.

The execution model: background, scheduled, and an approval gate

What actually changed in this expansion is when, where, and who is watching. Three axes move together. First, work keeps running in the background even when no active device is on. Second, scheduled runs, such as a 6 a.m. client prep, fire automatically at a set time. Third, at points that need human judgment the agent asks a clarifying question first, and any output requires the user's approval before it is published. If the first two axes push toward autonomy, the third pulls a brake on it. Putting autonomous execution and an approval gate in one product is the backbone of this announcement.

Why it matters now: the coding agent spills into the whole office

Cowork's web and mobile expansion shows the center of gravity of agent products shifting from coding to general knowledge work. According to Anthropic, over 90% of Cowork usage is non-coding knowledge work, and the largest categories are business operations and content creation. That covers tasks like reconciling finances, tracking contracts, and building presentations. This signals that the capability sharpened by a year or two of coding-agent competition is now spilling into office work broadly. As autonomous agents that began as developer tools descend into the daily work of finance, sales, and planning staff, the competitive front moves from code accuracy to how much work context you can safely delegate. The decision to widen to three screens reads in the same light. Office work does not happen only at a desk, so a session has to stay alive on the move for delegation to actually hold.

Between autonomy and control: the approval gate is the real design point

The notable part of Cowork's structure is not autonomy itself but the control placed on that autonomy. Background and scheduled execution turn the agent from a "chat window" into an "always-on worker," because work proceeds even while the user is not watching. The catch is that the less an autonomous run is watched, the greater the risk that bad judgments quietly accumulate. Clarifying questions and pre-publish approval are the buffer aimed exactly at this point. The agent stops and asks at forks where a human must decide, and anything leaving for the outside must pass through a human's hands. In the end, what this product sells is not "full automation" but "supervisable delegation." More than how autonomous it is, where it stops to call a human during an autonomous run is what decides the substance of trust.

What practitioners should check now

The first thing for a team to watch is the access order and timing. Cowork on web and mobile starts as a beta, opens first to Max tier users, and then widens to other plans over several weeks. Rather than planning around company-wide adoption today, it is safer to first validate with a limited group whether background and scheduled execution fit your real workflows. The next thing to watch is how the approval gate meshes with your organization's control rules. For work whose output leaves the company, such as financial reconciliation or contract tracking, the pre-publish approval step should connect to an approval chain so that delegation and audit requirements hold at once. The doubled usage benefit is extended through August 5, 2026, so using this window as a validation period to weigh cost against oversight burden is the core task of this expansion.

Source: Anthropic official announcement (2026-07-07, "Claude Cowork on web and mobile"); reporting by TechCrunch and SiliconANGLE.