OpenAI Launches Always-On Agent Dots Powered by gpt-6 astra at DevDay
OpenAI launched the always-on agent Dots at DevDay, powered by gpt-6 astra, equipped with a standalone cloud computer and browser, capable of working continuously towards user goals 24/7 and learning preferences from feedback.
Dots connects to over 4,000 applications through a plugin ecosystem, allowing conversations or voice calls in ChatGPT, Slack, and Teams. Users can open their cloud computer at any time to check progress and can authorize it to connect to their personal laptops for collaborative work. The background "active research" can only view connected applications in read-only mode and cannot send messages, change content, or control the browser and computer.
The first Dot is included in eligible Pro and Business Premium packages at no additional charge; enterprise, education, and healthcare workspaces require admin approval to enable trials. Sensitive operations such as changing passwords or permanently deleting data require explicit user consent. Conversations with Dot do not count towards ChatGPT usage limits, but tasks initiated by it in Codex or ChatGPT Work are still billed according to original rules.
Demonstration scenarios include: developers using dedicated Dots to monitor customer feedback and push bug fixes and new features; early testers' Dots drafting invoices after discovering missed billing, which are sent out after approval; calendars proactively suggesting takeout options and prices during overtime meals. The company claims that in the future, multiple Dots can collaborate, and a professional Dot with independent identity, IT-distributed hardware, and system integration will be previewed for enterprises.
OpenAI also launched the shared workspace ChatGPT Space, allowing colleagues, ChatGPT, and users' Dots to share the same project knowledge pool. By default, Dot can delegate coding tasks to Codex and has integrated with Microsoft’s Agent 365 for security controls to configure identity, credentials, and tools in enterprise environments.
In market mechanics, this is an event-driven product launch: paid subscription users first receive continuous online labor, while reasoning and cloud computer usage remain on OpenAI's account; Microsoft embeds enterprise security layers into the same channel through Agent 365. Beneficiaries are teams that have paid for high-tier packages and wish to offload monitoring, research, drafting, and bug fixing from human shifts; pressured are assistant products that still sell attention per conversation and competitors that have just rolled out consumer-grade agents in the form of independent cloud computers. Funding shifts from "per turn chat" to "per resident agent position," with applications accessed through plugins competing for the default invocation by Dot.
gpt-6 astra was previously positioned as a work-oriented flagship in computer usage, browsing, and engineering assessments; Dots transforms it from "only working when you open the window" to "running on its own machine even when the window is closed." Active research is restricted to read-only to separate long-term autonomy from write permissions, reducing the risk of system changes and message sending during unattended operation.
Source: Public Information
ABAB AI Insight
OpenAI's trajectory from the 2025 Computer-Using Agent and Operator has evolved from "giving the model a machine that can press buttons" to using ChatGPT Agent with virtual computers for multi-step tasks, and then to Codex integrating engineering into the same toolchain. Dots does not create a new model line but encapsulates the computer usage, browsing, and long tasks already trained in gpt-6 astra into a never-off identity: with a name, a cloud desktop, and a plugin ledger. Historically, the failure points of Operators and early Agents were that tasks would stop when prolonged or lose context when switching windows; Dots uses independent cloud computers and checkable desktops to change "not seeing what it's doing" into an auditable background workstation.
The capital path is clear: first, bind the resident agent to existing high-priced subscriptions with Pro and Business Premium, then redirect task execution back to Codex and ChatGPT Work billing, while cloud computers and reasoning remain on their own computing power. Partnering with Microsoft for Agent 365 integration hands over identity, credentials, and enterprise security controls to a distributor that has already occupied the office entry point, allowing Dots to access company systems rather than just personal chat boxes. The motivation is not to sell another round of conversations but to turn "workstations still running after hours" into a reason for renewal, while reserving IT procurement orders for future multi-Dot teams and professional identities.
The counterpart is Meta's Muse: both emphasize independent secure computers, continuous background work, and prior approval for sensitive actions. The difference lies in the entry point—Muse first targets consumer downloads and daily assistant mindsets, while Dots locks in paid work packages and Slack/Teams workstations. The industry phase has shifted from "which model answers better" to "who can reside in the enterprise permission system and be allowed to touch the system." Anthropic is pursuing another productization path for computer usage and coding agents, while OpenAI this time bundles flagship models, cloud computers, plugins, and office suites into a single colleague structure.
Structurally, this represents a technological substitution layered with industry chain reconstruction: what is being replaced is not a specific type of Q&A, but the duties of monitoring, following up, drafting, and cross-application transport that only progress when someone is at work. The mechanism strips writing permissions from active loops and moves the execution environment from the user's local machine to an isolated cloud computer, thus allowing long-term autonomy to be sold to enterprises in compliance. Pricing power shifts from per-token conversations to whoever controls the resident agent and the 4,000 application entries it can connect to by default.
ABAB News · Cognitive Laws
Intelligence that can be turned off is a tool; intelligence that cannot be turned off is an organization.
First give the machine an independent desk, then dare to hand it the keys.
Permissions are sold separately: read-only can reside, writing must knock on the door.