The conversation around generative artificial intelligence has moved beyond simple, text-based chat boxes. While early models focused on answering questions, drafting emails, and summarizing articles in a linear, prompt-by-prompt loop, the modern landscape is shifting toward autonomous agents. These systems do not just chat with users; they actively plan, execute multi-step workflows, and coordinate across local and cloud environments to deliver completed projects.
In a major step for this technology, Anthropic has officially expanded Claude Cowork—its specialized general-knowledge AI agent—from a desktop-only application to both web and mobile platforms.
The update is rolling out in beta, starting with subscribers on the high-tier Max plan, with plans to expand to Pro, Team, and Enterprise users over the coming weeks. By taking Claude Cowork beyond the desktop, the company is attempting to create a continuous, cross-device ecosystem.
A user can initiate a complex task on a desktop computer, walk away from their desk, close their laptop, and monitor the agent’s progress directly from an iOS or Android device. This structural evolution introduces a level of flexibility that could redefine remote work and administrative productivity.
Alongside this release, Anthropic published a detailed set of usage metrics compiled from 1.2 million anonymized Claude Cowork sessions across more than 600,000 organizations. The data challenges a common narrative in the tech sector.
While the industry has largely focused on AI coding agents and developer productivity tools, the real-world usage data shows that the vast majority of people use these agents to handle administrative, non-technical knowledge work. By building a mobile and web framework around these practical, daily workflows, Anthropic is positioning itself to lead the next phase of enterprise automation.
The Shift to Multi-Device Continuous Workflows
When Claude Cowork first launched, it was restricted to a desktop environment. This restriction existed because the agent required direct, local access to the user’s file systems, web browsers, and desktop tools to execute commands.
While this desktop-centric approach offered powerful automation capabilities, it came with a major operational drawback: the host computer had to remain turned on, awake, and connected to a stable internet connection for the agent to complete its work. If a user closed their laptop to commute or attend a meeting, the active work session would instantly freeze.
The expansion to web and mobile platforms addresses this limitation by transitioning the agent’s execution model to the cloud. When a user runs a remote session, the active computational work is handled entirely on Anthropic’s secure servers rather than on the user’s local machine.
This architectural change decouples the agent from any single physical device. It enables users to describe a task, point Claude to the necessary cloud resources or uploaded files, and close their laptops. The agent will continue executing the task in the background, entirely offline from the user’s perspective.
This continuous workflow structure allows for unprecedented cross-device collaboration. For instance, a customer success representative traveling between meetings can initiate a client-tracking dashboard on a laptop.
While they are in transit, the agent runs the necessary research, pulls data from connected spreadsheets, and formats the output. When the representative arrives at their destination, they can open the mobile app on their phone, review the finished deliverable, and approve it on the spot.
This level of seamless transition from desktop to mobile makes autonomous agents far more practical for remote, on-the-go professionals who cannot afford to sit at a desk waiting for an automation routine to finish.
Data-Driven Proof: Demystifying the AI Agent User Base
The launch of Claude Cowork on mobile and web is accompanied by usage data that reveals how employees are actually interacting with autonomous AI. The dominant narrative in the technology sector has long suggested that software engineers and developers are the primary beneficiaries of advanced AI assistants. Tools like GitHub Copilot and specialized coding agents have received the lion’s share of venture capital and media attention.
However, Anthropic’s analysis of 1.2 million anonymized Cowork sessions paints a starkly different picture. The data shows that a massive 91.3% of all sampled sessions involved tasks completely unrelated to writing or editing software code.
Only 8.7% of the total sessions were used for software development. Instead, non-technical knowledge workers have emerged as the driving force behind the platform’s adoption.
According to the data, business process and operations tasks represented the largest single category of usage, accounting for 33.4% of all sessions. These workflows include pulling scattered operational updates into a single report, building structured team onboarding checklists, and reconciling complex financial spreadsheets.
Content creation and marketing operations made up the next largest chunk of activity. This distribution proves that the actual market for agentic AI lies in the broad, diverse field of general knowledge work.
By bringing Claude Cowork to web and mobile, Anthropic is directly targeting this non-technical workforce. Employees in marketing, sales, human resources, and operations do not want to interact with command-line interfaces or manage local development environments.
They require simple, accessible interfaces where they can describe an outcome in plain English, upload their everyday files, and receive a completed spreadsheet, document, or presentation. Extending these features to the web and mobile apps lowers the barrier to entry, allowing non-technical workers to run complex automations from the same familiar interfaces they use for daily messaging.
Autonomous Core Features: Scheduled Tasks and Desktop Integration
The remote-execution model changes what Claude Cowork can do, particularly through the introduction of scheduled tasks. Under the previous desktop-only framework, recurring tasks were difficult to manage because they required the user’s computer to be active at the exact time the automation was scheduled to run.
Now, because work sessions can run remotely on Anthropic’s cloud infrastructure, users can configure Claude to run recurring, multi-step tasks on a precise schedule without needing any of their physical devices to be online.
For example, a remote team manager can schedule Claude to crawl specific partner websites every Monday morning, extract updated product pricing data, compare the findings to an existing internal inventory database, and format the results into an interactive CSV file.
By the time the manager opens their laptop to start the work week, the compiled spreadsheet is already waiting in their Claude account.
At the same time, the desktop version of Claude Cowork retains its deep integration capabilities. When running locally on a computer, the agent can access local directories, manage file structures, and interact directly with desktop software.
This dual-nature approach allows users to choose the best environment for their specific needs: using local execution for tasks that require manipulation of local system files, and shifting to remote cloud sessions for continuous, background workflows that span multiple devices.
The “Human-in-the-Loop” Push Notification Model
One of the persistent challenges of deploying autonomous AI agents in professional environments is the issue of trust. Allowing an algorithm to operate independently, manipulate files, and conduct research in the background can create anxiety for users who remain legally and professionally responsible for the final output. If an agent operates entirely out of sight, a small error or misinterpretation of a prompt can compound over a multi-step task, resulting in a completely unusable deliverable.
To manage this risk, Anthropic utilizes a highly structured “human-in-the-loop” oversight model. The fundamental rule of the platform is straightforward: nothing ships to third parties or gets finalized without explicit user review and approval.
When Claude Cowork runs a background session remotely, it constantly evaluates its own progress. If it encounters a highly ambiguous instruction, experiences a connector error, or completes the final draft of a deliverable, it halts the run and waits for feedback.
With the new mobile integration, this review loop is managed through push notifications. If Claude requires user input or a final sign-off, it sends a notification directly to the user’s smartphone.
The user can tap the notification, open the Claude app, view a step-by-step log of what the agent has done so far, and either provide clarification or tap “approve” to let the agent complete the task. This system transforms the employee’s role from an active, step-by-step tool operator into an executive manager who provides high-level guidance and final quality control, drastically reducing the cognitive overhead of managing complex automation routines.
The Cost, Access, and Promotion Timeline
Deploying and running autonomous agents remotely in the cloud is an incredibly compute-intensive process. Unlike a simple text model that generates a single response, an agent must continuously run code, analyze intermediate outputs, correct its own errors, and manage file streams over sessions that can last for hours.
To manage this massive computational load, Anthropic is rolling out the web and mobile beta updates in stages, prioritizing its highest-value users first.
The update is initially available to subscribers on the Max plan, which costs $100 per month and is designed for power users and advanced professionals who require high-frequency access and maximum usage limits.
Over the next several weeks, the beta will expand to users on the Pro plan ($20 per month), as well as Team and Enterprise organizational accounts.
To encourage widespread testing and gather deep user feedback during this critical beta phase, Anthropic is doubling Claude Cowork usage limits for all eligible subscribers through August 5. This promotional window allows enterprise teams to integrate the agent deeply into their daily workflows, test the boundaries of remote server execution, and evaluate the financial return on investment of delegating administrative tasks to an autonomous digital assistant.
The Security Paradox of Remote Agentic Access
While the expansion of Claude Cowork to web and mobile offers undeniable productivity benefits, it also introduces significant security risks. Giving an autonomous AI agent the ability to access local files, browse the web, and execute code remotely creates a highly complex attack surface that corporate IT departments must carefully evaluate.
The primary concern involves the threat of prompt injection and malicious data manipulation. If an agent is tasked with browsing a live website to conduct research or downloading public attachments from a vendor, it could encounter malicious text hidden on those web pages.
This hidden text can act as an injection attack, overriding the user’s original instructions and commanding the agent to perform unauthorized actions, such as exfiltrating sensitive data, deleting local files, or clicking on malicious links.
Because the agent operates in the background, a user might not notice this malicious deviation until the task is already complete.
To mitigate these risks, Anthropic runs all remote shell commands and code execution within highly isolated, sandboxed environments on its own servers. This design prevents any remote exploit from directly accessing the user’s physical machine or broader corporate network.
Additionally, the company has integrated strict network egress permissions, giving corporate IT administrators the ability to restrict the websites and cloud services that Claude can interact with.
Organizations can also completely disable web search capabilities for Cowork and Chat within their administrative dashboards. As autonomous helpers become a standard fixture of the modern office, maintaining a strict balance between agentic capability and zero-trust security architecture will remain a top priority for enterprise security teams.
Redefining the Role of the Modern Knowledge Worker
The expansion of Claude Cowork to the cloud, web, and mobile represents a maturity milestone for the generative artificial intelligence sector. By moving beyond the limits of the desktop and enabling background, cross-device execution, Anthropic is transforming AI from a basic, chat-based research assistant into a continuous, reliable administrative partner.
Furthermore, the real-world usage data released by the company proves that the true value of autonomous agents lies not in the niche field of software development, but in the massive, everyday world of general business processes, operations, and document management.
As these autonomous tools become more capable and easier to access, the role of the knowledge worker is undergoing a fundamental shift. Success in the modern remote office will increasingly rely less on the manual execution of repetitive digital tasks, such as copy-pasting data, formatting slides, and sorting files.
Instead, professionals must develop the skills required to direct, supervise, and refine the work of autonomous agents. By bridging the gap between desktop power and mobile convenience, tools like Claude Cowork are laying the groundwork for a more efficient, decentralized, and agent-driven future of work.
