OpenAI's Dots moves always-on agents from owned hardware into OpenAI's cloud

A Turing Post writer covering OpenAI's DevDay in San Francisco still runs an OpenClaw agent on a physical computer, one the writer can close ^[17][3]. In that setup it is basically free, and the writer plans to keep it ^[3]. "I appreciate having that choice," the writer wrote ^[4].

During the OpenClaw rush, people bought Mac minis so the agent could stay online, because it stops working when it goes offline ^[5]. A San Francisco ride-share driver told the writer, "I barely ever use a laptop now. All my agents are on my phone." ^[6] The pitch for Dots assumes users want foresight: an OpenAI employee told Turing Post that Dots resemble Meta's Muse but are much more proactive, working out what might be needed before anyone assigns it ^[16].

Dots take over the upkeep. An agent needs files, tools, permissions and somewhere to run. It also needs enough saved context that coming back to a project does not mean explaining it all again ^[11]. OpenAI maintains much of that for Dots ^[11]. With a self-hosted agent, all of it stays with the team.

For now the team does not get to pick the model. Asked about routing, OpenAI's Tibo said: "We're working on routing. We didn't announce routing today, but this is something that we're working on." ^[9] Until that ships, the cheaper Sol 6.1 model announced at the same event cannot pick up routine Dot work ^[1]. Turing Post also reported that OpenAI postponed Astra 6.1, its next model ^[7].

Cost is the part a planner cannot fill in yet. Chat with a Dot does not count as tokens ^[10], so the number of conversations tells a team nothing about spend. A better test is whether a task finished without anyone having to brief the agent again ^[11]. The Turing Post account does not include which plans get Dots or what unattended work costs.

Of the AWS and Microsoft arrangements, the writer wrote: "It wants a substantial role in your work even when you use someone else's software." ^[13][14]

I think a team piloting agents in Slack or Teams should start with hosted Dots ^[2]. The upkeep list is the work a small team least wants to own. The tradeoff is that every task runs on Astra until routing ships, and the team no longer owns a machine it can switch off ^[8][1].

Two tests sort the rest. The first is whether the work has to start with nobody present, triggered by an event in a connected app ^[12]. The second is whether someone on the team must be able to shut the machine down personally. Hosted Dots fit unattended work with no shutdown requirement. When unattended work does need the shutdown, self-hosting fits, with a named owner for the files, permissions and saved context. For attended work without the requirement, an ordinary chat agent covers it and the always-on computer adds little. Attended work that needs the shutdown belongs on a local agent run on demand, which is the setup the Turing Post writer kept ^[3].