AI agents capable of working after users leave have entered a new phase beyond the early competition, with OpenAI moving through Dots, Meta through Muse, and xAI through Grok Bot, while Anthropic chose to bring Cowork into the Claude app rather than offer it as a standalone product with a separate name and marketing identity.
Cowork enables scheduled tasks to run after a user’s computer is shut down, and it began operating this way in July. After being integrated into Claude in September, it initially became available to Pro and Max subscribers, with plans to make it available to Team subscribers and free users. Anthropic has also had Managed Agents for long-running hosted agents since April, and launched Claude Tag within Slack in June, with its own service identity and audit log.
A Race in Which Ideas Are Repeated Quickly
The article explains that the market for persistent agents is still so new that it is impossible to speak of an established winner. Projects such as Clawdbot emerged; it began as a project for a system running on an additional computer and receiving commands through WhatsApp or Telegram, then became Moltbot and later OpenClaw. In less than a year, ideas such as advanced research, terminal coding agents, profiles, an agent’s own operational identity, and working inside the chat window have moved among competing labs.
Accordingly, launching a similar feature does not necessarily indicate a strategic delay. For Anthropic, integrating Cowork into Claude may be a way to introduce the agent quickly to a user base already building with Claude, rather than spending effort creating a new brand. The article argues that users who pay for subscriptions and repeatedly use the agent matter more than a splashy launch that achieves large downloads and then loses momentum.
Different Models for Reaching Users
Muse achieved rapid reach; Apptopia estimated its number of daily users in the United States on iOS during its first 12 days at 359,000, compared with 231,000 for the ChatGPT app during the same period, while Sensor Tower estimated downloads at more than 3.4 million through September 24, with the companies’ estimates differing. More than 95% of Muse users are also connected to Facebook, according to Apptopia, giving Meta a broad distribution channel. Muse offers a free tier and paid plans priced at $20 and $100 per month, in addition to connectors with Shopify and Stripe that allow agents to purchase products.
By contrast, OpenAI and xAI started from the paid end. Dots requires a ChatGPT Pro subscription costing between $100 and $500 per month, or a Business Premium or Enterprise account, while Grok Bot reached 418,000 weekly users in mid-September, according to Bloomberg. These companies are betting on subscriber retention to determine whether usage will continue after the novelty of the launch wears off, rather than on a mass download base alone.
What Changes in Practice?
Anthropic’s approach gives the agent a place close to developers and teams already using Claude. The article says that an internal version of Claude Tag is responsible for about 65% of code changes among product teams, according to Cat Wu. It also presents examples of agents being used to simulate supply chains, conduct overnight quality tests, inspect invoices, track broken releases, and, in some cases, merge fixes.
But Claude does not yet exist in the same consumer context as Muse within Meta-owned WhatsApp. Users need to open the Claude app or mention the agent inside Slack. This difference may matter to the average user, while a technical team may care more about access scope, identity, and the operations log. The article distinguishes between an agent that shares a single cloud computer and a single set of credentials, as with Grok Bots, and an agent with an independent service identity and permissions restricted by channels, as with Claude Tag.
Trust Is the Real Test
A persistent agent holds credentials and can perform actions while its owner is absent, so technical capability alone is not enough. xAI’s documentation warns against treating Grok Bots as a security barrier, while the OpenClaw Foundation says that most information technology departments prohibit agent platforms. Anthropic’s default settings tend toward caution; Claude requests permission before execution unless the user changes that, and Claude Tag maintains an audit log.
Editorial reading: This competition reveals that the distinguishing advantage is no longer simply having an agent that operates automatically, because ideas move among companies within weeks or months. The practical difference will be determined by how easily the agent can be integrated into the work environment, the limits of its permissions, the reviewability of its actions, and whether completing a task requires more supervision than carrying it out manually. The article has not yet established that Cowork or any competitor has settled these questions; the products are still early, and the most important test is whether users return to entrust them with repetitive and sensitive work.