OpenAI is moving beyond the familiar chatbot model with Dots, a new kind of persistent AI agent designed to keep working after an employee closes the chat window — monitoring projects, using software, responding to changing information and bringing completed work back for approval.
Announced at OpenAI’s DevDay 2026, Dots run on the company’s flagship GPT-6 Astra model and receive their own cloud computer and browser. It will be accessible in ChatGPT on the mobile app and website, and OpenAI co-founder and CEO Sam Altman said during his DevDay keynote that it will soon be available on other messaging apps and even over the phone as an audio model, as well.
Dots can connect through OpenAI’s plugin ecosystem to more than 4,000 applications, communicate with users through ChatGPT, Slack and Microsoft Teams, and gradually learn an individual’s preferences, standards and working habits.
OpenAI is pairing those individual agents with ChatGPT Space, a new collaborative layer where employees, ChatGPT, Codex and Dots can work against the same shared context. Space replaces ChatGPT’s existing Library for Pro, Business and Enterprise users and acts as a home for Pages, files, presentations, spreadsheets and other team materials.
Rather than keeping AI work confined to an employee’s private chat, teams can bring people and agents into shared documents, tag a Dot or ChatGPT for help, and continue building on work already completed by colleagues or other agents.
That combination makes Dots less like another ChatGPT feature and more like OpenAI’s attempt to create a persistent digital coworker: software that employees can delegate responsibility to rather than simply ask questions of.
For enterprises, the distinction matters. Much of the first wave of generative AI adoption has required employees to repeatedly initiate interactions: open an assistant, supply context, request an output, review it, then start again when the situation changes.
Dots are designed around continuity instead. OpenAI says a dot can keep several projects moving simultaneously, accept new work without forcing users into separate conversational threads and operate while its human counterpart is doing something else.
From copilots to employees delegating entire workflows
OpenAI’s examples make clear where it wants enterprises to deploy the technology first.
For developers, a dot can watch incoming customer feedback, identify recurring requests or bugs, scope smaller fixes, build and test them, and return completed pull requests accompanied by videos demonstrating the changes. The developer can continue working on a major feature while the agent handles smaller maintenance jobs.
For product and marketing teams, the agent can learn an organization’s audience, positioning and creative standards, then revise launch material when the underlying product changes.
OpenAI also shows Dots updating scientific analyses as new experimental evidence arrives, revising figures and explanations when findings change. In another example, a sales-focused dot compares an enterprise prospect’s requirements against product documentation and account history, identifies remaining technical tests, constructs a proof of concept, suggests an appropriate solutions engineer and keeps the proposal up to date as the deal changes.
Content teams get a similar treatment: a dot can ingest an interview transcript, find potential clips, prepare show notes, draft social posts and carry subsequent editorial changes across the associated materials.
ChatGPT Space extends that model from individual task execution into persistent team artifacts. OpenAI says Pages can remain synchronized with connected tools, letting teams create documents that update as underlying information changes.
A project page could, for example, pull action items from Slack, email and calendars while ChatGPT or a Dot keeps owners, deadlines and next steps current as priorities shift. Teams can also tell ChatGPT what sources to check, what parts of a Page to update and how often to run the workflow.
Dots also work natively across files in Space. OpenAI says an employee can interact with the same Dot through ChatGPT, Slack or Microsoft Teams and ask it to keep a shared Page current as plans change. That makes Space potentially important to the broader Dots model: the agent can follow an employee across communications channels while maintaining work that remains visible to the wider team.
The common thread is not generation, but workflow ownership.
An employee does not merely ask the AI to produce one artifact. The employee establishes an objective and standards, while the agent monitors the underlying work and responds as circumstances change. In theory, that could shift some knowledge workers from executing every intermediate step toward supervising a portfolio of work performed partly by agents.
The addition of Space pushes that idea another step: from one employee supervising one agent toward groups of employees and agents operating against shared documents, project plans and organizational knowledge. Instead of passing static AI-generated files between colleagues, OpenAI is pitching shared work products that humans and agents can continue revising together.
That could increase output without necessarily requiring enterprises to completely redesign their existing software stacks. Dots can operate through browsers and connected applications rather than requiring every business process to be rebuilt as an AI-native application.
But it also means companies are no longer deciding merely whether an AI can read an internal document. They are deciding whether it can continuously observe operational systems, manipulate software and make decisions about what to do next.
An agent that can work when nobody is babysitting
OpenAI calls one of Dots’ background capabilities “proactive research.”
When a user is not actively working with the agent, a dot can inspect information in already connected applications and look for things that may require attention. OpenAI places an important restriction around that background activity: the tools used for proactive research are read-only and cannot send messages, modify content or control the user’s computer or browser.
Once actual changes or external actions enter the picture, permissions become more complicated.
Organizations and users select which applications a dot can access through ChatGPT’s existing app controls. OpenAI is also introducing Custom Rules, which can permit particular actions, require approval or prohibit them. An Activity View lets users inspect background work and intervene.
An additional auto-review system checks potentially consequential actions against the user’s instructions, Custom Rules and OpenAI’s built-in safety requirements before determining whether the work can proceed autonomously or requires approval. Some sensitive operations, including changing passwords, remain reserved for the human user.
Each dot operates inside its own cloud computer, separated from the employee’s machine unless the user explicitly connects the two. OpenAI says supported websites can use saved credentials without exposing the underlying password to the model. Dots can nevertheless be granted permission to connect directly to a laptop or other device, potentially giving them much deeper reach into an employee’s working environment.
That makes governance central to the product rather than an enterprise feature that can be added later. A mistake made by a chatbot may produce an inaccurate paragraph. A mistake made by an agent operating authenticated business software can potentially alter records, send information or trigger downstream workflows.
ChatGPT Space adds another governance dimension because the outputs of those agents can become shared organizational artifacts. OpenAI says workspace administrators’ existing sharing rules continue to apply, and teams can grant view or edit access to Pages and later change or revoke it. Files uploaded directly into a Page inherit that Page’s permissions, while merely linking to an external file does not grant collaborators access to the original — although any material copied or summarized onto the shared Page becomes visible to everyone who can access it.
OpenAI also draws a boundary between a user’s private ChatGPT context and shared Space content. Sharing a Page does not expose private conversations or personal memory to coworkers. But if ChatGPT uses something from a person’s memory and writes it onto the Page, that information becomes visible to collaborators, making the Page itself an important point for employees and administrators to review what information crosses from private AI context into shared work.
OpenAI acknowledges that Dots can still make mistakes and advises users to review consequential work.
For Business, Enterprise and Edu customers, OpenAI says workspace content is not used to improve its models by default. Personal-plan users can choose whether Dot conversations and work are used for model improvement. The company says it does not train directly on a dot’s proactive research or its private notes to itself.
OpenAI's ChatGPT Space documentation reiterates that OpenAI does not train on Business or Enterprise workspace data by default.
It also introduces a subtle consideration for mixed personal collaboration: each participant’s own training and memory settings apply when their ChatGPT works with shared content, and OpenAI warns that material someone shares could potentially affect another collaborator’s memory or be eligible for training under that collaborator’s individual settings.
Specialist Dots could turn the concept into an organizational labor layer
The more ambitious enterprise play comes from something OpenAI is only beginning to pilot: specialist Dots.
A personal dot acts for one user. Specialist Dots instead receive their own organizational identity, credentials and access to company systems and are assigned a defined business responsibility.
OpenAI says it has already experimented internally with agents working in procurement, invoice processing, email marketing, customer support and commercial contracting. It is beginning focused pilots with enterprises in which OpenAI engineers work directly with customers to define each agent’s responsibilities, permitted tools and approval processes.
OpenAI is also working with Microsoft to integrate those specialist agents into Microsoft Agent 365, allowing businesses to govern them through security and management systems they may already use.
If that model works, enterprises may eventually treat agents less like software licenses assigned to humans and more like new machine identities that themselves require onboarding, credentials, access policies, monitoring and offboarding.
Space provides a potentially important human-facing counterpart to that machine-identity layer. Rather than requiring employees to follow an agent into a separate administrative console, OpenAI is designing shared Pages and project areas where people can see the artifacts agents are maintaining, leave comments, make edits and @mention ChatGPT or a Dot to take the next step. OpenAI says multiple employees with edit access can work on the same Page simultaneously, with each person using their own ChatGPT against that common document.
That is potentially a major change for IT departments. The employee directory of the future may have to account not only for people and service accounts, but persistent AI workers operating between them.
Dots joins a rapidly escalating battle for the personal AI agent
OpenAI is not entering an empty market.
Over the past several weeks, a cluster of companies has begun converging on a remarkably similar idea: an AI assistant with a persistent identity, its own computing environment, access to a user’s accounts and applications, memory of what matters to that person, and enough autonomy to act without continuous prompting.
The most conspicuous example is Meta’s Muse, launched Sept. 8. Meta describes Muse as a personal agent running inside a dedicated Muse Secure VM, with its own browser and the ability to work across connected apps, remember details about users and proactively help them accomplish goals.
Consumers appear to be responding. TechCrunch reported on Sensor Tower data estimating that Muse had already surpassed 3.4 million downloads by Sept. 24, less than three weeks after launch. Other analytics providers have produced different estimates — Apptopia put it at 4.3 million while Appfigures estimated roughly 2.3 million — but all point to unusually rapid initial adoption. Sensor Tower also estimated 2.8 million downloads in Muse’s first two weeks.
Muse reached the top of both Apple’s and Google’s U.S. app-store rankings during this surge.
Meta is already trying to carry that momentum into work. On Tuesday, it expanded Muse with small-business integrations spanning tools including Shopify, QuickBooks, Stripe and Canva. The Wall Street Journal reported that about one-third of Muse users have connected a business-related account, while more than 1,500 businesses have applied for Muse integrations.
Startups are chasing the same interface. Instinct describes its assistant as a persistent personal agent connected to email, messaging, a user’s screen, audio, location and devices. Users can call or text it, and the company pitches it as software that follows up on neglected threads, arranges transportation and books services without forcing the user into a new interface.
Manus, meanwhile, introduced Manus 2.0 this week with a new cloud computer, automations and Cue, which it describes as a new application for personal agents.
SpaceXAI’s Grok Bot, launched in August, similarly gives agents their own computers, allows them to sign into existing apps, keeps them running around the clock and asks users for approval when necessary.
Open-source OpenClaw, which helped popularize agents that live inside existing messaging services and control software on a user’s behalf, represents another branch of the same movement.
Together these launches suggest that persistent personal agents are becoming a distinct product category — and potentially the next major interface battle in AI.
The competitive question is shifting from whose model gives the best answer to which agent a person or business is willing to trust with continuous access to the machinery of everyday life and work.
That makes personality and familiarity strategically important, not merely cosmetic. A named assistant that remembers preferences, messages like a colleague and develops continuity over months may become harder to replace than a generic chatbot — especially once it accumulates permissions, workflows and knowledge about how its user operates.
For enterprises, that also creates a looming boundary problem: employees may arrive with increasingly capable personal agents just as companies are deploying sanctioned corporate ones.
Dots is also the connective tissue for a much larger OpenAI push
A flurry of OpenAI announcements at DevDay 2026
OpenAI used DevDay to announce more than 20 products and updates, including GPT-6.1 Sol, new high-speed model options, additional privacy infrastructure, computer use in the Agents API, expanded plugins, event-triggered automations, collaborative ChatGPT workspaces and deeper Slack and Teams integration.
ChatGPT Space may be one of the clearest examples of how those pieces fit together. OpenAI describes it as the shared home for Pages and files created, uploaded and shared in ChatGPT, replacing the previous Library for eligible users. Pages combine text, images, charts and interactive tools in documents that employees and AI can edit together in real time, while connected tools such as Google Drive and Slack can supply additional context. Collaborative slides and spreadsheets are also coming.
For enterprises, ChatGPT Space could turn Dots from individual assistants into participants in shared team workflows. Teams can bring goals, plans, designs and other material into one common environment, then ask ChatGPT or a Dot to maintain dashboards, translate a teammate’s design into engineering tasks, flag blockers or keep project documents synchronized with new information.
Recurring tasks can be delegated to automations that gather information and act on schedules or external events. In practice, that creates a model where employees and multiple AI agents can work against the same project context, hand work back and forth, and update common deliverables rather than operating as isolated one-to-one assistants.
Viewed together, those products give Dots much of the infrastructure a persistent enterprise agent needs: models, computers, tools, event triggers, communications channels, shared workspaces and organizational controls.
The first dot is included at no additional charge for Pro and Business Premium customers, although OpenAI has not disclosed the numerical allowance for intensive work, the eventual cost of additional Dots, pricing for higher speed or workload capacity, or commercial pricing for specialist Dots. Enterprise, Edu and Healthcare customers can try the beta when an administrator enables it.
Those missing numbers will matter. If persistent agents genuinely take over chunks of administrative, engineering, sales, research and creative work, enterprises will eventually need to measure them less like chat subscriptions and more like labor and infrastructure: how much useful work each agent completes, how often a human must intervene, what mistakes cost, and whether the productivity gains justify the compute and governance overhead.
Dots makes that future substantially less theoretical.