OpenAI on Tuesday unveiled Dots, persistent AI agents powered by GPT-6 Astra that continue working even when users step away, marking a shift from traditional AI assistants that wait for each new prompt. Announced at the company's DevDay event, Dots run on dedicated cloud computers, operate their own browsers, and connect to more than 4,000 applications through OpenAI's plugin ecosystem. Unlike conventional agents, a Dot can manage multiple projects simultaneously and transfer its knowledge across ChatGPT, Slack, and Microsoft Teams.
The agents can handle substantial portions of development workflows, from spotting problems to delivering solutions. According to OpenAI's developer demonstration, a Dot can monitor customer feedback for repeated requests, outline minor enhancements and bug repairs, construct and test modifications, and submit completed pull requests with video documentation of the changes. This lets developers concentrate on bigger features while reviewing the agent's finished work. Within OpenAI itself, Dots begin investigating when bugs surface in Slack, and in another scenario, a Dot can convert fresh design files into functional applications while teams concentrate on customer input. The work extends beyond code: OpenAI demonstrated Dots rerunning scientific analyses as fresh data arrives, revising sales proposals when customer needs shift, and transforming interview recordings into video clips, show notes, and social media content.
OpenAI separates finding work from acting on it to manage the risks of an always-on, unsupervised agent. When users aren't actively collaborating with a Dot, it performs what the company calls "proactive research," scanning connected applications for opportunities to assist, but that access is read-only—the Dot cannot dispatch messages, alter app content, or take control of a browser or computer. Actions that might impact user accounts or share information move through a distinct "auto-review" step that evaluates them against user instructions, OpenAI's safety requirements, and any Custom Rules the user has established. Those rules can permit certain actions to proceed autonomously, demand approval, or block them entirely, while particularly sensitive tasks like password changes always remain under user control. The company says Dots include protections against malicious instructions—a genuine concern for an agent reading content from thousands of connected apps—alongside a monitoring system that can pause or halt a Dot when it identifies a safety issue. Users can track all Dot activity, including background work, through an Activity View and redirect the agent from there. OpenAI cautions that Dots can still make mistakes and recommends reviewing work with significant consequences.
OpenAI also previewed specialist Dots, which organizations provision with dedicated identities, credentials, IT-supplied hardware, and access to systems of record so each owns a specific workflow rather than supporting a single employee. The company has tested this approach internally in procurement, invoice processing, email marketing, customer support, and commercial contracting, and plans to start external deployments through enterprise pilots where OpenAI engineers collaborate with customers to define each Dot's responsibilities, tool access, and review points. Dots are rolling out now to ChatGPT Pro and Business Premium users in eligible markets, with one Dot included in those plans at no additional charge, while Enterprise customers, including Education and Healthcare, can access the beta once a workspace administrator enables it. Conversations with a Dot don't count toward ChatGPT usage limits, though tasks it initiates or manages in Codex or ChatGPT Work draw from those products' limits as usual. Plans include an allowance for deeper work, with extended limits for the first month after launch, and OpenAI plans to let customers add more Dots and increase either a Dot's speed or the volume of work it can handle each month. For developers juggling competing priorities, the ability to delegate routine but time-consuming tasks to an agent that persists across platforms could reshape how engineering teams allocate attention. Organizations will need to weigh the efficiency gains against the governance overhead of managing agents with broad system access and the cultural shift of trusting AI to complete work without human oversight at every step.