OpenAI launches Dots, agents that run in the background and deliver ready-made pull requests

Each Dot runs on its own computer in the cloud, tests code, opens pull requests with video, and keeps working on the project without anyone needing to watch the conversation.

OpenAI presented Dots at DevDay 2026, a class of agent the company describes as "always on": it takes on a project, works for hours or days without supervision, and comes back with an update, a question, or a finished deliverable. The announcement was reported by B9, which also drew the most obvious parallel: Meta had presented Muse, with an almost identical proposal, three weeks earlier, on September 8.

For those who build automation and integrate systems, Dots is interesting less for its polished chat and more for the architecture behind it. According to the description OpenAI gave of the product, reproduced by B9, each Dot runs on its own virtual computer, with a dedicated browser, is powered by GPT-6 Astra, and connects to more than 4,000 apps through OpenAI's plugin ecosystem. Interaction happens through ChatGPT, Slack, Microsoft Teams, and voice calls, with iMessage and RCS in testing.

The example that speaks the language of developers

Among the use cases OpenAI itself highlighted, according to B9's report on the announcement, one is clearly aimed at engineering teams: a Dot can track user feedback, implement small fixes, test the generated code, and deliver pull requests already accompanied by video showing the change working.

The detail that matters here isn't the fix itself, it's the format of the delivery. The Dot doesn't hand you a snippet of code to paste: it opens the PR, runs the tests, and visually documents what changed. The human comes in afterward, to review and approve, not to orchestrate each step manually.

This changes the role of whoever receives that PR. Review stops being "reading a diff you yourself requested" and becomes "auditing a decision another agent made on its own," with the video serving as evidence that the expected behavior still holds.

Specialist Dots: the part built for production

Besides the personal Dot, OpenAI described the so-called specialist Dots: agents that companies will be able to configure with a role, identity, credentials, and access to specific systems. According to the statement cited by B9, this model is already being tested internally in areas such as procurement, invoice processing, email marketing, support, and contracts.

This is the piece that comes closest to real systems integration, not a chat copilot. A specialist Dot with its own credentials and access to a billing system is, in practice, an automation service with identity and permissions, only driven by natural language instead of business rules written in code.

The trade-off is the usual one with production automation: you gain configuration speed (describing the task instead of programming the flow), but you lose predictability. A deterministic script fails the same way every time; a generative agent can interpret the same request differently in different runs.

The permission problem Meta already exposed

The autonomy question got a concrete example even before Dots launched. According to a report reproduced by B9, a user of Meta's Muse had their home address mistakenly shared with a buyer on Facebook Marketplace, without explicit authorization for that action. Meta revised how it presents permissions after the episode.

OpenAI devoted a good part of the Dots presentation to this same point, according to B9's report. Per that description, the user chooses which apps the agent can access and can set rules that allow, block, or require approval for specific actions. There is also an automatic review before operations that could affect accounts or expose information, and some tasks, like changing a password, would remain mandatorily with the person.

In short: the permission model described has three layers: a rule declared by the user (allow, block, require approval), automatic review before sensitive actions, and a total block for a small set of critical operations. It's a granular authorization architecture trying to compensate for the fact that the agent makes decisions that used to require an explicit human click.

For those who already deal with IAM, RBAC, and access policies in traditional systems, the design sounds familiar. The difference is that here the "action" isn't a predictable API call: it's what a language model decided to do to fulfill an objective described in free text, which makes after-the-fact auditing as important as the prior rule.

What changes for those who build automation today

The basic unit of generative AI until now was the conversation: request, response, next request. With Dots and Muse, the session can last for days, the system observes changes in the project, and it even suggests work before any new request.

This has direct implications for anyone who currently handles long-running tasks with a cron job, webhook, message queue, or an orchestrator like Airflow. A Dot promises to replace part of that pipeline with a natural-language description, but it trades execution predictability for scope flexibility, and that isn't a free gain in any system that needs strong auditing or bit-for-bit determinism.

Availability also imposes practical limits for now, according to what B9 reported about the announcement:

• Dots would arrive on the ChatGPT Pro and Business Premium plans, with a beta version for Enterprise
• The first Dot would be included in those plans, but heavier workloads would have their own quotas
• Each user could create only one Dot for now; more Dots or more capacity are expected to become an additional charge in the future
• OpenAI's stated goal for later is to have entire teams of Dots working together

This limit of one agent per user, today, is what separates the announcement from real production automation. A CI pipeline that depends on a single agent shared by the whole team has bottleneck and queueing built in, which suggests that the current version of Dots better serves individual, supervised tasks than replacing an entire team's automation infrastructure.

Where it isn't worth it, for now

The video PR case is appealing, but it's worth remembering that it still depends on human review before the merge, exactly like any PR from an external contributor. Automation that requires a guarantee of behavior (deterministic rollback, auditable compliance, response SLA) shouldn't swap its current pipeline for an agent that, in OpenAI's own words, can still make mistakes.

The real gain shows up in tasks that already tolerate asynchronous supervision: tracking feedback, preparing drafts, keeping project documentation up to date as scope changes. It's draft automation with review, not a replacement for critical infrastructure, and that's how it's worth testing before putting any Dot in charge of a production system.