OpenAI has introduced Dots, a new feature that transforms AI assistants from reactive chatbots into persistent agents capable of working autonomously in the cloud. Launched on September 29, Dots are designed to continue executing tasks even after a conversation ends, marking a significant shift in how AI can be integrated into daily workflows.
Each Dot operates within its own dedicated cloud computer and browser environment, isolated from the user's personal device. This architecture allows the agent to pause and resume work as needed, manage multiple delegated tasks, and maintain continuity across sessions. The system leverages multiple state layers, including conversation history, ChatGPT memory, and saved notes about user preferences, to ensure coherent and context-aware performance.
Key Features and Architecture
The core innovation lies in the persistence of these agents. Unlike traditional chatbots that wait for user prompts, Dots can proactively track projects, wake themselves to execute scheduled tasks, and use connected tools to perform research, data analysis, document preparation, and even software development. They can also delegate subtasks to background agents, ChatGPT Work, or Codex, enabling complex, multi-step workflows.
Access to Dots is being rolled out gradually, initially available to Pro 100, Pro 200, and Pro 500 subscribers over the age of 18, excluding users in the European Economic Area, the United Kingdom, and Switzerland. Business Premium and Enterprise access is expanding worldwide, with enterprise administrators required to enable the feature manually as it defaults to off. Mobile access follows desktop setup and depends on a supporting app update.
Safety and Approval Mechanisms
Given the increased autonomy, safety is a critical component. OpenAI has implemented an automatic review system that evaluates actions that could affect accounts or share information. This review checks instructions, app permissions, custom rules, and built-in safety requirements, and can either let an action proceed, request user approval, or mandate a manual step for sensitive operations like password changes.
Research access is intentionally limited: proactive research tools can read connected information but cannot send messages, edit app content, or control a computer. Any follow-up action requires an additional permission check, highlighting the separation between reading and acting, which carry different levels of risk.
Custom Rules and Limitations
Users can establish custom rules for recurring tasks, choosing from options such as act without asking, act after an explicit request, ask before acting, or hand the step back. However, these rules are instructions, not guarantees, and OpenAI's controls guide warns that Dots can make mistakes. Users must also be mindful of stop controls: pausing a Dot halts its main task but does not automatically stop delegated jobs or future schedules, which must be reviewed and stopped separately.
Market Context and Implications
The launch of Dots comes amid growing competition in the AI agent space, with major tech companies and startups alike racing to deliver more autonomous AI solutions. The ability for agents to operate independently in the cloud could reshape productivity tools, potentially impacting workflows in finance, research, and software development. However, the lack of public data on task completion rates, error rates, and uptime means that real-world reliability remains unproven.
An earlier peer-reviewed study of 30 deployed AI agents found documented incidents or security concerns in eight, with prompt-injection vulnerabilities identified in two of five browser agents. While Dots was not part of that dataset, the findings underscore the importance of robust safety measures in persistent agents.
What to Watch
As the rollout progresses, key milestones to observe include whether eligible users receive access smoothly, how well context is preserved across mobile, Slack, and Teams interactions without leakage, and whether administrators can effectively monitor activity records and approval prompts. These observable deployment metrics will be more telling than promised intelligence gains, as the industry watches to see if persistent agents can deliver on their potential without compromising security or user trust.
