AI GUIDEPartner content

OpenAI DOTS: the architecture of 24/7 background AI agents, virtual instances, and the reality of digital employees

A technical analysis of the OpenAI DOTS architecture: how background AI agents with a virtual environment, browser, and access permissions in Slack and Microsoft Teams are built, as well as the hidden security and limit-related problems.

Affiliate link: your price stays the same and the project earns a commission.

The concept of personal assistants in chat is gradually reaching its ceiling. Users still have to sit in front of a screen, write prompts, wait for generation to finish, and manually transfer the result between services. At the DevDay 2026 conference, OpenAI showed an attempt to change this interaction model by introducing a system DOTS — autonomous background agents that operate continuously in their own cloud containers.

Agent anatomy: a virtual computer instead of a conventional API call

Unlike classic GPTs, which were essentially layers over a system prompt with several API integrations, each DOT receives a dedicated environment:

  • An isolated cloud instance (micro-VM) with its own file system;
  • An autonomous headless browser capable of authenticating on supported websites through a secure password store;
  • The flagship GPT-6 Astra base model as the core for planning and decision-making;
  • An interface for connecting to more than 4 000 enterprise services through an updated plugin protocol.

This architecture changes the workflow. The agent no longer requires a human to be continuously present in the web interface. The user formulates a goal, after which the agent switches to background execution: searching documentation, cloning repositories, running the build, checking logs, and creating a pull request. If a decision point arises during the process that requires human input, the agent sends a structured notification to Slack, Teams, or the ChatGPT mobile app.

Context synchronization between Slack, Teams, and ChatGPT

One of the main practical problems with agent systems has always been memory fragmentation. If a conversation began in chat while details surfaced in the team's workplace chat, the model lost the thread of its reasoning. In the DOTS architecture, this issue is resolved through a shared Workspace context layer.

A developer can set a basic task by voice through a smartphone on the way to the office. Later, when a colleague clarifies the technical requirements in a Slack thread, the DOT joins the discussion, reads the new input, and adjusts the plan without a repeat briefing. At the same time, the user retains full control: the interface provides desktop streaming for the agent's virtual machine, displaying open browser tabs, terminal commands being executed, and the file tree in real time.

Specialized roles: how configurations are structured

Instead of a universal assistant that tries to do everything at once, OpenAI is betting on role profiles with strictly defined tools and access policies:

Profile Primary tool stack Key workflow scenario
Iggy (QA/CI Engineer) Git, GitHub API, test runners, log parsers Tracking bug reports, localizing errors in code, running regression tests, and submitting PRs
Flip (Product Release Manager) Jira, Linear, Google Docs, Notion, Slack Monitoring pre-release checklists, synchronizing documentation, and distributing changelogs
Data/Green (Data Researcher) Python, Jupyter, SQL connectors, BI systems Recalculating analytical models in the background when data marts are updated, and searching for anomalies
Alfred (B2B Presales) Salesforce, HubSpot, PDF readers, API sandboxes Analyzing the customer's specifications, checking compatibility with the product, and preparing a rough prototype
Jojah (Content Producer) Audio/video transcribers, Markdown editors Analyzing source materials, extracting key points, and drafting publications

Enterprise security: Agent 365 and digital identity

Releasing autonomous entities into internal corporate networks creates enormous risks of leaks and uncontrolled changes. In partnership with Microsoft, the company is implementing integration with Microsoft 365 and Entra ID.

An enterprise agent is registered as a separate digital entity with its own service account. At the policy level, three action boundaries are distinguished:

  1. Autonomous operations: reading repositories, compiling code in a temporary environment, running tests, and searching documentation.
  2. Operations requiring confirmation: sending external emails, pushing to protected repository branches, conducting financial transactions, and making changes to production databases.
  3. Absolutely blocked actions: exporting private keys, changing passwords, and managing other users' permissions.

The economics and hidden costs of background inference

The main question in deploying such systems is not technological but economic: token economics. A background agent independently working through a complex problem for 4–6 hours can consume several million tokens through iterative attempts, context reading, and error debugging.

Under the standard Pro ($200/month) and Business subscriptions, the company includes one basic agent; however, quotas for demanding background computing tasks through Codex are strictly enforced. Real-world industrial use of teams of such agents will inevitably require either switching to expensive enterprise contracts or tightly controlling the depth of reasoning (reasoning effort) for each subtask.

Compare models before you start

The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.

Browse models

Affiliate link: your price stays the same and the project earns a commission.

SEO Mind42 editorial team

We explore SEO and neural networks in practice: test services on our own projects, verify prices and limits against primary sources, and share things you can put to use the same day.

📚 Reference guide to SEO and AI 🔄 Materials are updated 🕐 Updated: 3 October 2026

Related reading

All in this section →