OpenAI Dots put persistent agents inside ChatGPT

OpenAI has introduced Dots, a type of ChatGPT agent that can continue working after a conversation ends. Powered by GPT-6 Astra, each dot receives a persistent cloud computer, browser, memory, and access to approved apps. It can monitor ongoing work, launch tasks, and return results for review without waiting for another prompt.

The release extends ChatGPT from request-and-response sessions into long-running software workers. A dot remains available across supported interfaces and can pursue user-defined goals in the background, subject to permissions and approval checks.

One agent across several channels

Each dot has a customizable name, avatar, and handle such as @yourname-dot. OpenAI says it can connect to thousands of apps, retain project context, run scheduled tasks, and use its cloud browser to complete web-based work.

The same dot can receive messages in ChatGPT, Slack, and Microsoft Teams, as well as through voice calls. Its memory follows it across those surfaces, allowing a conversation started in Slack to inform later work in the ChatGPT desktop app.

A dot can also divide work among background agents, open parallel cloud threads, and create Codex tasks on a connected computer. For development work, it can operate in a Codex cloud environment linked to a repository. When a website blocks cloud browsers or requires authentication, the dot can request a private sign-in flow or return browser control through a Take over button.

The work Dots can absorb

OpenAI’s initial examples focus on recurring development and operations work that spans several tools:

• Triaging bug reports and feature requests from connected apps.
• Investigating failing builds and scoping improvements against team priorities.
• Building and testing changes with Codex, then returning completed pull requests for review.
• Monitoring several projects and carrying relevant context between supported channels.

Users can inspect a dot’s cloud computer while it works or grant access to a local machine. Local access requires a separate permission from Codex and Work Sync, and each account can link one personal computer at a time.

Work continues between messages

A dot tracks task progress, determines when work should pause or resume, and decides when user input is required. During idle periods, it can search approved sources with read-only access for updates relevant to its assigned goals.

Actions that modify an account, send information, or change external systems require the appropriate integration permission or explicit approval. The operating model separates proactive research from state-changing actions that carry greater risk.

Memory draws from three sources: the current conversation, the user’s ChatGPT memory, and private notes maintained by the dot. Those notes can contain preferences, prior decisions, and unresolved tasks. Because they persist across channels, instructions given in one interface can affect later behavior elsewhere.

Actions pass through approval checks

Before a dot changes an account or shares information, an automated review determines whether it may proceed, must request approval, or needs to hand control to the user. Custom rules can require draft review before messages are sent, although those rules cannot override OpenAI’s built-in safety requirements.

Ahead of the conference where it introduced Dots, OpenAI said it had canceled plans to debut GPT-6.1 Astra, citing elevated risk after recent incidents involving autonomous hacking. Dots use the existing GPT-6 Astra model. OpenAI also released GPT-6.1 Sol, a lower-cost model described as approaching Astra’s capability.

Paid tiers get the first rollout

Access depends on the user’s plan, region, device, and workspace settings:

Conversations with a dot do not consume ChatGPT usage allowances. Tasks launched through Work or Codex draw from those products’ quotas. OpenAI says eligible plan allowances will not be charged for usage during the month following launch. Bloomberg also reported that OpenAI introduced a $500 Pro tier alongside the release.

Persistence changes the agent architecture

Most agent frameworks, including products built with the Agents API, represent work as bounded runs: a request starts, tools execute, and the run ends. Dots add a long-lived agent object whose conversations serve as entry points into the same workspace. Its browser sessions, files, memory, and task state can survive across interactions.

For developers, that architecture changes the operational requirements. Persistent per-user virtual machines raise questions about idle compute costs, tenant isolation, credential storage, software patching, audit logs, cancellation, and recovery after partial failures. Cross-channel memory also requires clear controls for retention, deletion, and access boundaries.

Dots place OpenAI in the growing market for persistent assistants, including Meta’s Muse. The immediate competitive focus is practical work across existing tools, such as responding to a bug reported in Slack, preparing a pull request, or sending an overlooked invoice after the required approval.

Production limits remain

Teams connecting a dot to production systems need to account for several documented constraints:

1. Stopping a task does not reverse actions that have already completed.
2. Dots can make mistakes even when approval rules are enabled.
3. Some websites block cloud browsers, requiring access through a connected local machine.
4. Pausing the main task does not cancel delegated background work or scheduled runs.
5. Local access works only while the ChatGPT app is open and the linked computer is online.

For teams building agent products, Dots provide a concrete implementation of a cloud-resident agent with a dedicated workspace, cross-channel memory, app integrations, parallel execution, and approval controls. Evaluating the product will depend on how reliably those components handle long-running work, partial failures, sensitive credentials, and human review.