Anthropic's Cowork merge puts a scheduled background agent inside the main Claude app

OpenAI's own example shows how a Dot starts work. A bug alert lands in Slack, and the Dot begins digging on its own ^[5]. Cowork is described in terms of schedules. According to Matt Burns's newsletter at The New Stack, it runs scheduled jobs after the user closes the laptop ^[7]. Both run in the cloud ^[4][1].

A scheduled job fires at a set time and fetches what it was told to fetch. An event-driven agent fires when an upstream system changes, so its value depends on which systems it can watch. Connector reach is the second thing to compare. The newsletter reports a count for Dots ^[4]. It does not report one for Cowork, or say whether Cowork can start work from an alert the way OpenAI's example does ^[7].

Connector reach also depends on the far side. Meta's Muse has connectors to Shopify and Stripe checkout that let its agents buy things ^[11]. Amazon started blocking Muse two weeks after launch, according to the newsletter ^[12].

The cloud machine per agent sets OpenAI's price. Sam Altman called Dots a premium product because each one needs so much compute, Burns reports ^[10]. Meta took the other route with Muse, offering a free tier and paid plans at $20 and $100 a month ^[11].

Self-hosting is the older model, and its record is short. Peter Steinberger pushed Clawdbot to GitHub last November as a weekend project ^[13]. It ran on a spare machine, accepted instructions through WhatsApp or Telegram, and used Claude for most of its work ^[13]. Anthropic sent lawyers in January, and the project was renamed Moltbot, then renamed again to OpenClaw a couple of days later ^[19]. Security became a headache for the project, and in February Steinberger joined OpenAI ^[14].

The newsletter's one head-to-head result is Jessica Wachtel's test ^[8]. Burns cites it next to Cowork and OpenAI's earlier Work mode, and Dots only launched on Tuesday ^[2][3]. For the result to transfer, a team's recurring jobs would have to resemble those three developer tasks.

Burns wrote that "Always-on agents are too young to have a winner, and the moats are shallow" ^[16]. Anthropic shipped its version without a cute name or a fuzzy mascot, in his words ^[18]. A platform team gets one fewer name for its vendor register.

I think the choice follows from the jobs a team already has. A nightly report or a weekly sync fits a scheduled agent inside the Claude app the team already uses ^[7]. Work that starts with an alert fits the event-driven design OpenAI has documented ^[5].