OpenAI's Dots pairs always-on agents with their own cloud computers and app access

OpenAI has launched Dots, a product that packages persistent AI agents — each with its own cloud computer, browser, memory and app access — into a single system. The details come from a dev.to breakdown of OpenAI's official announcement and launch video, which argues that the interesting part is not any one feature but the way familiar agent capabilities are stitched together into an always-on whole.

Always-on by design

The core shift is the interaction model. A conventional assistant works in request-response: you ask, it answers, the exchange ends. Dots inverts that. As the dev.to author describes it, you give a Dot a goal and it keeps working in the background, returning when it has something useful. OpenAI characterises these agents as staying on around the clock, pursuing goals and juggling several projects at once. The mental model moves from chatting with a model to delegating a job.

A computer of its own

Each Dot is provisioned with its own cloud computer and browser, giving the agent an actual environment to work in rather than just a chat window. It can use that browser, connected apps and files, and users can open the Dot's computer and watch what it is doing. That visibility matters. The architecture becomes model plus computer plus tools, and the browser turns into a working instrument — visiting sites, gathering information, comparing options and preparing results — rather than a demo feature.

Memory that carries over

Dots are designed not to reset after every task. According to the dev.to summary, OpenAI says they pick up your goals, preferences, standards and sense of what quality means to you through ongoing interaction and feedback. The intended loop is that feedback teaches preferences, later tasks need less explanation, and output gradually starts to resemble your own work. That is what makes the approach potentially valuable for extended work like coding, research, writing and content production, where context normally evaporates between sessions.

Chat, Slack, Teams or voice

Dots are not tied to a single interface. They can be reached through ChatGPT, Slack and Teams, and OpenAI also lets you place voice calls to your Dot. The notable part is which layer is persistent: the Dot, not the channel, is the entity that carries context. You can start a task in ChatGPT, continue it in Slack and follow up by voice, with the interfaces acting as windows onto the same underlying agent.

Apps and actions

Through its plugin ecosystem, Dots can reportedly connect to more than 4,000 apps. That is what pushes the system from reasoning toward execution — an AI that discusses a task versus one that can carry it out, with permissions and approvals sitting inside the stack as a first-class part of the design.

Specialist Dots ahead

OpenAI is already signalling a next step: specialist Dots for organisations, where individual agents hold their own identity, credentials, tools and responsibilities. The dev.to post sketches the resulting trajectory — one assistant, then one persistent agent, then multiple specialised agents, then effectively a team of agents. The author also notes that the architecture described is a personal simplification based on OpenAI's public materials, not an official diagram.

Why it matters

None of the individual ingredients here is unprecedented: persistent agents, sandboxed execution environments, long-term memory, browser automation and tool plugins have all shipped in some form before. The significance is the packaging — a major vendor combining all of them into one product, with the unit of usage shifting from the conversation to the delegated job. If Dots works as described, it changes how work gets assigned, monitored and reviewed, since an agent that runs unsupervised in the background raises questions of reliability, oversight and cost. The dev.to breakdown's honest conclusion is worth keeping in mind: the bet is on integration, not invention, and whether that bet pays off depends on execution details the announcement does not fully address.