Anthropic Embeds Agent Capabilities Into Claude as Rivals Race for Always-On AI Adoption

The competition for always-on agents has intensified rapidly. OpenAI unveiled Dots at DevDay this week, autonomous agents with dedicated cloud infrastructure and browser access to over 4,000 apps connected to ChatGPT. Meta's Muse arrived earlier and has already surpassed ChatGPT's mobile installation benchmarks. xAI shipped Grok Bot in August. Anthropic's response came two weeks ago as an update to Claude, integrating Cowork—a feature that has run scheduled jobs since July—directly into the main application, without the marketing fanfare of a separate product name or mascot.

The stakes differ by company. Meta requires Muse to become a mainstream success; its Reality Labs division has invested tens of billions without delivering a breakout consumer product, and Llama 4's weak reception prompted a reorganization around a new superintelligence lab. Muse reached an estimated 359,000 daily U.S. iOS users in its first 12 days—exceeding ChatGPT's 231,000 at the same stage—and Sensor Tower counted more than 3.4 million downloads by September 24. More than 95 percent of Muse users also use Facebook, giving Meta a built-in distribution advantage. The platform offers free and paid tiers at $20 and $100 monthly, plus integrations with Shopify and Stripe that let agents execute purchases.

OpenAI and xAI have taken a different path, targeting paying customers from the start. Dots requires a ChatGPT Pro subscription ($100 to $500 monthly) or a Business Premium or Enterprise account; Sam Altman described Dots as a premium product due to its computational demands. Grok Bot reached 418,000 weekly users by mid-September, according to Bloomberg, and xAI launched Team Bots on Monday. Neither company depends on consumer scale to validate the concept; retention among existing paid users will reveal whether agents retain engagement once the initial appeal fades.

Always-on agents are too early to have a winner

The current wave of always-on personal agents emerged less than a year ago. Peter Steinberger released a weekend project called Clawdbot to GitHub in November, a system that ran on a spare computer and accepted prompts via WhatsApp or Telegram using Claude. Anthropic's legal team intervened in January, leading to a rename to Moltbot and then OpenClaw days later. Security concerns mounted, and in February Steinberger joined OpenAI. This week, the foundation overseeing the project released an early, pre-1.0 version of OpenClaw Enterprise for internal company testing.

The pace of feature adoption across competitors has accelerated dramatically. Meta's Nat Friedman acknowledged that Muse drew heavy inspiration from OpenClaw, and users examining Muse's code discovered a SOUL.md personality file nearly identical to OpenClaw's. Anthropic's September integration of Cowork into Claude followed OpenAI's Work mode by two months—though Cowork's cloud and scheduled-task capabilities shipped on July 7, two days before Work mode launched. When Jessica Wachtel tested both systems across three developer tasks, she found them tied on accuracy, with ChatGPT delivering faster results and Claude providing more detailed analysis. At this stage of product maturity, a comparable feature launch allows a lab to observe how its own user base applies the technology.

Meta needs Muse to be huge. The others need paying users.

Muse has achieved significant traction. Apptopia's analysis of iOS data alone showed 359,000 daily U.S. users during Muse's opening 12 days, surpassing ChatGPT's 231,000 at the equivalent stage. Sensor Tower's estimate of more than 3.4 million downloads as of September 24 demonstrates strong early adoption, though different measurement firms report varying figures. Meta's dependency on this success is substantial. Reality Labs has expended tens of billions without producing a mainstream consumer offering. Llama 4's disappointing reception last year prompted Zuckerberg to restructure his AI organization around a new superintelligence lab, notably acquiring a 49 percent stake in Scale AI and recruiting its founder Alexandr Wang.

Muse represents the first significant product from that restructuring. Its growth has been fueled by Meta's existing user base: Apptopia determined that over 95 percent of Muse users maintain Facebook accounts. The platform operates on a freemium model with paid subscriptions at $20 and $100 monthly, and includes integrations with Shopify and Stripe checkout that enable agents to complete purchases. However, this scale-first strategy carries complications. Amazon began blocking Muse two weeks after its launch, introducing a new competitive dynamic in e-commerce.

Anthropic shipped its version to the people already building with Claude

Anthropic's strategy reflects a focus on its existing user base. The company released the updated Claude quickly, distributed it to paying subscribers, and embedded the agent capability within the application those users already operate. The merged Claude reached Pro and Max customers immediately, with Team and Free plans arriving soon. Claude Tag, a proactive Slack teammate with its own identity and audit trail, has been available to teams since June. Cat Wu noted that an internal version accounts for approximately 65 percent of product teams' code changes. Developers have had access to Managed Agents for hosted, long-running agents since April.

Anthropic has not pursued Muse's distribution model. While Muse operates within Meta-owned WhatsApp, Claude still requires users to open the Claude application or mention it in Slack. This workflow distinction may matter for general consumers. Teams integrating an agent into their codebase prioritize what systems the agent can access. The architectural differences carry security implications: Grok Bots share a single cloud computer and one set of credentials across the entire roster, whereas Claude Tag joins a Slack workspace with its own service identity and channel-scoped permissions. Organizations considering agent deployment must understand these distinctions before granting agents access to credentials.

Anthropic should target the builders already experimenting with always-on agents. Writers on Towards Data Science have documented throughout the year what developers accomplish with these systems. Samir Saci deployed a team of OpenClaw agents on a supply chain simulation to identify late shipments. Eivind Kjosbakken's guide to operating a fleet of OpenClaw bots describes nightly quality assurance bots that test applications and report defects, plus agents that validate invoices. Both examples ran agents on OpenAI's Codex. OpenAI operates its own OpenClaw agent, Androidclaw, which identifies broken builds and, in certain cases, applies fixes, according to VentureBeat. Anthropic requires this category of work running on Claude.

Builders, coders and developers prove essential to product development because they articulate their questions openly in GitHub issues and Discord threads, and those answers shape the product roadmap. Trust remains the obvious obstacle. An always-on agent holds credentials and operates without human oversight. xAI's documentation explicitly warns users not to treat Grok Bots as a security boundary, and the OpenClaw Foundation reports that most IT departments prohibit agent platforms entirely. Anthropic's design defaults toward caution: Claude requests permission before acting unless instructed otherwise, and Tag maintains its own audit trail. If builders do not return, or if each completed task demands more supervision than manual execution, the experiment has failed. However, if they continue discovering useful work to delegate, their questions and insights become the product direction.

The meaningful competition for Claude, Dots and Muse centers on becoming the agent platform that users entrust with the next job.