OpenAI Dots Tutorial: Codex Agent Workflow Guide

Introduction

OpenAI released Dots on September 29, 2026. It is an always-on persistent agent powered by GPT-6 Astra, running on cloud infrastructure with its own built-in browser and compute environment. Dots can maintain persistent responsibilities, operate continuously, and resume tasks without requiring users to return to a desktop session. On October 10, Day 5 updates expanded compatibility to iOS and Android mobile ChatGPT clients, enabling end-to-end Dots workflow creation directly on mobile devices. On the same release cycle, OpenAI rolled out the beta version of Composer Q Predictions for Codex. This feature anticipates the next user message after a conversation, presenting suggested continuations with the Tab key for user confirmation before execution.

This tutorial covers the full workflow for Dots. It explains how to confirm access eligibility, create a dot agent, connect plugins and local computing resources, spawn both local and cloud Codex workloads, inspect task progress in Activity panels, pause and terminate running jobs, and understand the underlying logic of Composer Predictions.

1. Dots Availability: Verifying Access and Rollout Boundaries

Dots does not ship as a standalone downloadable application. To access Dots, users install or update the ChatGPT application, then check their account dashboard for the Dots entry point. As of October 10, 2026, official documentation defines three supported creation interfaces: ChatGPT mobile apps, ChatGPT desktop application, and desktop web browsers. Mobile web pages are not supported for dot creation, and mobile users must install the latest native ChatGPT app.

The Day 5 update released on October 10 extended complete dot lifecycle support to iOS and Android. Users with eligible plans can directly access Dots after updating their ChatGPT client, and the full workflow can be completed on mobile without desktop browser access.

Dots rollout proceeds in phases, with tiered availability:

• Pro 100, Pro 200, Pro 500: Available for users aged 18+ in European economic zones, the United Kingdom and Switzerland.
• Business Premium: Global phased rollout.
• Enterprise: Globally available but disabled by default; workspace administrators must manually enable the feature.

Even for users who meet plan, age and geographic criteria, the Dots entry may not appear in accounts during the staged rollout. If Dots is missing, follow this troubleshooting sequence:

1. Update ChatGPT mobile or desktop application, or log into the same account via desktop browser.
2. Verify subscription tier, age, geographic region and workspace permissions.
3. For enterprise accounts, confirm the workspace admin has activated Dots.
4. If all prerequisites are satisfied but the entry remains hidden, wait for phased feature rollout. Never download unofficial Dots installers from third-party websites.

Important billing note: Chat dialogues within Dots do not consume standard ChatGPT quota. However, ChatGPT Work and Codex tasks initiated or managed by a dot follow their separate usage billing rules.

2. Creating a Dot: Three Required Configuration Steps

Creating a dot only requires three simple setup steps. Linking a personal computer is optional. After the Day 5 update, all three steps can be finished on iOS, Android and desktop clients.

1. Open Dots within ChatGPT mobile, desktop or web interface, and complete the product introduction walkthrough.
2. Assign a name and visual appearance to the dot; default values can be retained and modified later.
3. Select plugins for the dot to access, such as email, calendar and file services. This step can be skipped temporarily.

Once configured, a single dot can resume ongoing tasks across different devices. ChatGPT, Slack, Microsoft Teams and voice calling are separate communication channels connected to the same agent. Switching channels does not reset the dot’s memory state.

It is critical to understand that communication channels, plugins and personal computer access are three independent permission groups. Adding a dot to a Slack channel does not automatically grant it access to cloud storage or local computers. When connecting Gmail, Google Drive, GitHub and other plugins, the dot inherits permissions from the authenticated user account.

3. Running Your First Task: Authoring Persistent Responsibilities

Dots is optimized for long-running, evolving responsibilities, rather than one-off question-and-answer prompts. When defining the first task, clearly specify deliverables, reference materials, approval boundaries and notification rules. The agent must understand what requires continuous updates, and which events trigger user intervention.

A usable starting prompt template is shown below:
> Maintain this project continuously. Review the provided specification documents, meeting transcripts and pending task list. Keep track of current objectives, assignees, deadlines, blockers and decision items. Update the list whenever source materials change. Notify me only when deadlines face risk, requirements conflict, or my approval is required. Before sending external messages, deleting content or modifying permissions, obtain my prior consent. Start with a small initial summary for review, and do not expand scope without approval.

This prompt contains five core components: target deliverables, approved data sources, change tracking scope, user confirmation triggers, and notification rules. After the first output, users can refine classification logic and grant additional sources or automated permissions incrementally.

For recurring scheduled workflows, define start time, end time and recurrence rules, and request the dot to restate the plan. Example:
> For the next four weeks, inspect this checklist every workday at 9:00 AM China Standard Time. Alert me only if deadline risks emerge, and confirm the schedule before execution.

4. Connect Plugins, Messaging Channels and Local Computers

Dots includes a cloud-native browser and runtime environment. It can research information, process cloud-hosted files and execute software without connecting to a local PC. Local computer access is only required when tasks need local files, native desktop applications, browser session context or local Codex execution.

The workflow to enable local computer access:

1. Open ChatGPT desktop and navigate to your dot profile.
2. Locate the Computers section and select *Your computer*, then choose Allow access.
3. Review permission disclosures and confirm by selecting Allow access again.

As of October 10, 2026, one dot can connect to only one personal computer at a time. The host machine must remain online, and the ChatGPT desktop application must stay open during task execution. An Offline status indicator only means temporary disconnection, and it does not revoke existing permissions. Users can revoke access manually via the Revoke access option.

Local PC permissions, Codex computer linking and Work Sync are independent authorization scopes. Enabling one permission group does not activate the others, and the dot cannot automatically read all ChatGPT workspace data.

5. Spawning Codex Tasks: Local versus Cloud Environment Selection

Dots can launch new local Codex jobs, continue existing local Codex sessions, or provision cloud-based coding workloads within pre-published Codex Cloud environments. The primary distinction between the two modes lies in where code and tooling execute.

New Codex tasks only receive context provided by the dot for that specific job. They cannot automatically access the full conversation history between the user and dot. When spawning tasks, users must explicitly define project paths, objectives, constraints, validation criteria and delivery formats.

5.1 Spawning Local Codex Tasks: Keep Your Computer Online

Local mode fits scenarios where code has not been pushed to remote repositories, dependencies are installed locally, or tasks require native desktop resources. First complete the local computer connection workflow, then describe the task in the dot chat interface.

Template for creating a new local Codex task:
> Create a new Codex task on my connected computer, opening [absolute project path]. Read AGENTS.md, README and existing test suites within this project, then resolve [specific issue]. Do not alter unrelated files, push changes remotely or create new branches. After finishing, run relevant tests, and provide a summary, test outcomes, unresolved risks and items requiring my decisions.

Template to resume an existing local Codex task:
> Continue the Codex task named [task name or unique identifier] on my computer. Inspect its current modifications and test results, then adjust according to [specific changes]. Preserve existing work and avoid reinitializing the project. If the task cannot be uniquely identified, prompt me for clarification.

Once a local task is created, users can open the corresponding Codex thread on the connected PC. Switching the selected computer for the dot does not migrate existing tasks; subsequent instructions will still target the original machine.

5.2 Spawning Codex Cloud Tasks: Publish Environment First

Cloud mode requires pre-provisioning a Codex Cloud environment. This process packages repositories, dependencies and configurations into reusable task foundations. An environment only needs to be created once, and dots can launch multiple coding jobs against it.

Official environment creation workflow:

1. Within a new Codex task, select Work in > Cloud.
2. Open Select environment and choose Create environment.
3. Pick the GitHub repository to clone; authorize GitHub connection if not already linked.
4. Select Get started to let Codex scan the repository, install dependencies and validate tooling.
5. Fill in missing permissions, version specifications, command or service information.
6. Review the setup report, resolve outstanding issues, save and select Publish.
7. After seeing Environment published notification, the environment becomes available for new tasks.

Users can also start environment creation via Settings > Codex Cloud > Environments > Create environment. Saving a draft does not publish it; dots cannot use environments marked Unpublished.

After publishing, use this prompt template for dot:
> Using my published [environment name] Codex Cloud environment, create a new coding task targeting [repository name]. Implement [specific requirement]. Read repository documentation and test specifications, apply modifications and run associated test suites. Prepare deliverable binaries, but avoid unapproved service deployments. Escalate requests for new permissions, secrets or product decisions to me. After completion, summarize file changes, test results, remaining risks and recommended next steps.

Cloud environment tasks can continue running even when the personal computer is asleep or offline. Network access rules only define which addresses the environment can reach; they do not automatically supply service credentials. Repository connections, network access and secrets must be configured separately.

6. Day 5 Release: Three Major Improvements for Dots and Codex Collaboration

The October 10 Day 5 update introduced three functional upgrades for dot-to-Codex task spawning, addressing historical pain points including lost background context and unintentional task fragmentation.

6.1 Historical Context Retrieval

Dots can now search ChatGPT conversation logs and extract context from prior Codex threads and automated jobs. When a requirement was discussed in earlier dialogue, the dot can automatically pull relevant background information into Codex, reducing redundant re-explanation.

6.2 More Accurate Thread Boundary Judgement

The agent is better at deciding whether to continue work within an existing thread or spawn a separate Codex task. It analyzes the scope of work and priority rules, preventing unrelated work from being split into disconnected threads. This improves continuity for multi-stage engineering workflows.

6.3 Inspection and Scheduled Task Management

Dots can view and modify scheduled tasks within ChatGPT workflows. Users can check upcoming plans, pause periodic jobs or adjust execution timelines directly through conversation prompts, without navigating to the dedicated Scheduled panel.

After modifications, dots continue handling pull request collection and target delivery. Users can add additional requirements in the chat at any time. Once a dot submits a task to Codex, it tracks progress until completion.

7. Inspecting, Querying and Resuming Tasks in Activity

After dot jobs start, the Activity panel acts as the central audit entry. Open the dot profile, select Activity, then enter individual tasks to review edits, file changes, results, failures and timestamped records. Users can approve, reject or supply manual input within this view.

When validating Codex outcomes, verify these four items at minimum:

1. Whether modifications match original objectives and avoid altering unrelated files.
2. Whether test suites execute successfully and capture failure cases.
3. Whether secrets, environment variables and access credentials remain properly isolated.
4. Whether new risks, service deployments or production changes require formal approval.

Additional instructions can be sent directly within active tasks, or users can request the dot to inspect outputs and continue coordination. Tasks created by a dot retain their original host computer or cloud environment; switching the dot’s connection target does not automatically migrate running workloads.

8. Pause, Activity and Scheduled Task Termination Rules

Task stopping mechanisms vary by task category. Pause only suspends the current active task of a dot; it will not automatically terminate delegated Codex workloads or future scheduled jobs.

• Primary dot tasks: Use Pause, then Resume to continue later.
• Delegated sub-tasks: Open the corresponding task in Activity and terminate independently.
• Scheduled recurring tasks: Disable or delete them inside Scheduled management panel.

Stopping a task does not roll back completed file edits, message delivery or other external side effects. For work handling sensitive data, review Activity and Scheduled records first, then verify final state in external connected systems.

9. Conclusion

Dots provides a complete workflow for persistent AI agents. Users create dots via ChatGPT on iOS, Android or desktop, define persistent bounded responsibilities, optionally connect local computers, spawn local or cloud Codex workloads, and monitor execution in Activity panels. The October 10 Day 5 update optimizes context retrieval, thread segmentation and scheduled task controls for dot-Codex collaboration. Composer Predictions further streamlines interaction by suggesting next actions with Tab-triggered confirmation.

OpenAI’s official documentation confirms three core layers: Dots persistent agents, Tasks and memory management, and Codex Cloud environments. As of October 10, 2026, Dots remains in phased rollout, and account access, subscription rules and workspace policies are subject to ongoing adjustments. Production usage should always refer to live in-platform prompts and official documentation.

When building production AI agent pipelines, developers often need unified routing and authentication for multiple model endpoints. An API gateway simplifies orchestration across heterogeneous model services.

International access: https://4sapi.com
Domestic access: https://4sapi.org