OpenAI Dots: Better Always-On Agents That Continue Working After You Log Off

DevDay became the stage for something other than a feature update.

OpenAI introduced Dots, always-on autonomous AI agents that keep operating long after a chat window closes. Altman framed them as digital assistants with dedicated cloud environments, browsers, and integrations reaching across more than 4,000 applications through OpenAI’s plugin ecosystem, subject to the user’s existing permissions. Slack, Teams, or voice: each agent carries a distinct identity and learned preferences, with a mandate to persist.

The launch’s edge is timing. OpenAI added Dots-specific information to the GPT-6 Astra system card on September 29, 2026. Autonomous capability is accelerating far faster than the frameworks built to govern it.

Persistence, and not chat, is the new interface, and the transition toward continuous digital labor has begun.

What Dots Actually Do Differently

ChatGPT already retains memory, supports agentic workflows, and handles scheduled tasks, but a dot is designed to take ongoing responsibility for a project, proactively review connected information, and keep making progress between conversations.

OpenAI’s “dots,” by contrast, introduce an architectural shift, where they hold long-term responsibility for ongoing projects.

Long-running workflows that continue across conversations fall within their purview.

A developer’s dot might monitor feedback, engineer fixes, test code, and submit pull requests with video demonstrations; for scientists, the same mechanism re-runs analyses automatically as new data arrives.

However, by tracking background tasks, such as drafting an overdue invoice for user approval, Dots transform AI from a reactive chatbot into a proactive, enduring collaborator.

Dots run on their own cloud machines. During proactive research – when the user is not actively working with the dot – background tools are restricted to read-only access. Other authorized or scheduled work can take actions subject to normal permissions and approval rules.

Custom Rules give users four behaviors for any given action: take action without asking, take action if pre-approved, ask before acting, or hand off entirely to the user.

Password changes and transferring money between financial accounts fall outside that spectrum entirely – those stay with the user, always.

Persistent Agents Raise New Authorization Questions

Persistent agents carry a safety problem that resists easy resolution: how does a system confirm that permission granted yesterday still holds tomorrow?

OpenAI testing exposes the friction between proactive evaluation and definitive security. OpenAI ran 100 simulated rollouts, each containing 500 emails, for 50,000 total emails – 16,600 malicious emails. The company reported zero scored attack successes in that specific bulk prompt-injection test.

As tasks accumulated, moderate scope violations increased. OpenAI reports that Astra’s chain of thought is less monitorable than GPT-6 Sol’s under some conditions, including adversarial tests where the model knows it is being monitored. OpenAI therefore says monitoring cannot rely on chain-of-thought alone.

OpenAI’s dots-specific testing is company-published, but Astra’s broader system card includes external work from organizations such as the UK AI Security Institute and Apollo Research. Real-world deployment evidence remains limited.

Disconnecting an app stops future access, but information already absorbed into a dot’s context is not automatically removed. The dot’s context can persist until you delete or reset it.

Who Gets Access and What It Costs

OpenAI has drawn a line, and most of its users sit on the wrong side of it. Dots, the company’s new always-on agent, is available only to upper-tier subscribers. Free and Plus users do not have Dots at launch, though OpenAI says it plans to expand access.

Entry begins with Pro 100 at $100 monthly or Pro 200 at $200 monthly, excluding the European Economic Area, Switzerland, and the UK for Pro users. Business Premium access spans supported ChatGPT regions. Enterprise, Edu, and Healthcare can access the beta when an administrator enables it, though it is initially turned off by default.

OpenAI also introduced Pro 500 at $500 per month, a broader ChatGPT subscription with the highest included usage allowance, and Astra Ultrafast, which is not a Dots-specific tier.

Some Hacker News commenters criticized the paid-plan requirement. Critics pointed to Meta’s Muse, which launched without a paywall, and asked why OpenAI chose to gate a capability with obvious practical value.

OpenAI’s Dots and Muse converge on persistent, cross-session AI architecture rather than distinct consumer branding. Delegating complex, long-horizon workflows ultimately hinges on trusting an evolving, unproven safety framework.