Delegate to Dot, Collab with Codex: A New Collaborative Workflow for Developers

Dominik Kundel from OpenAI's Developer Experience team had already invested considerable effort configuring Codex: using memory to track preferences, installing plugins, setting up automations, and even dedicating a "chief of staff" thread to coordinate work. Yet after using Dots for a while, he found the most noticeable change was that he no longer had to keep every single follow-up step in his head.

Being able to write code does not mean taking care of every follow-up task for you. The author's previous workflow was:

Bringing an Appshot (application screen snapshot) or a Slack message link to Codex, asking it to implement features, and instructing it to monitor deployment status and reply back in Slack with the results; but with every new requirement, he had to separately ask it to set up automations to track feature or release schedule changes, configuring everything repeatedly each time.

The real burden lay in having to organize instructions and configure follow-up methods separately for every new task.

Now, the author can simply @mention his Dot directly within the same Slack thread. Chopper takes on the task, resolves the issue, replies with the results back in the original thread, and keeps an eye on release-related changes to alert him. This connects delegation, execution, reporting, and follow-up into a unified loop, so the author does not have to stitch this workflow together manually every time a new task arrives.

The emergence of Dots is designed precisely to take over this burden of organization and follow-up. It is not meant to replace Codex, but rather is built on the same open-source Codex Agent execution harness (the runtime system that enables models to call tools and drive tasks forward), responsible for taking on tasks and continuously following up.

This also explains why developers who have already configured Codex can still benefit from it: a Dot can inherit the connected ChatGPT / Codex plugins on your account, using them to search for information, work across existing tools and services, and leverage existing relevant context. You do not need to redo your entire prior configuration just to add a delegation entry point.

Dot Is a Personal Agent for Continuous Delegation

Put simply, Dot is an always-on personal assistant powered by the GPT-6 Astra model. When the author wrote his article, only one Dot could be created, but you can give it a custom name, a personalized appearance, or even use a photo of your pet as its avatar.

For example, Dominik Kundel named his Dot Chopper and customized an avatar featuring an orange creature wearing a small black hat.

The defining characteristic of Dot is persistent presence and cross-surface coordination. It has its own dedicated Handle (such as @tibo-alfred or @yourname-dot), allowing you to delegate tasks directly to it without having to pick a specific Codex thread before getting started.

From Switching Chat Windows to a New Way of Delegating

Moving from using Codex directly to collaborating with Dot is reflected primarily in three key shifts:

1. Cross-Channel Continuity: Converse Anywhere, Context Never Resets to Zero

When communicating with coworkers or family in daily life, we might switch between messaging apps, phone calls, and social media, yet the background context of our discussion persists without needing to re-explain everything from scratch every time. Dot achieves something very similar:

• Continuous handoffs across contexts: A colleague reports a bug in Slack, and the author forwards it to his Dot (Chopper), which begins searching for the necessary information and starts working; on his evening commute, the author can call Chopper directly on the phone to keep discussing this bug or explore an entirely different new topic; back at his desk, the author can continue the conversation in the desktop chat or directly review the work it completed. Throughout the entire process, he is interacting with the very same Dot. The same Dot can leverage relevant context across channels, but messages between Slack and ChatGPT are not automatically copied to each other. Newly delegated tasks also receive only the context necessary to complete that task, rather than automatically inheriting all conversation history; sharing private chat information with others still requires your explicit authorization.
• Simultaneous multi-modal concurrency: While on a phone call with Dot, both parties can communicate via text at the same time. For example, Chopper can send content during the call for the author to review, and the author can type responses or send links on screen without hanging up (this is how the author iterated on articles in ChatGPT Space); while the call continues, Dot can also send an alert when another task finishes or proactively notify you if an urgent matter arises.
• Memory and note-taking mechanisms: In addition to using existing ChatGPT memory and related conversations, Dot features an independent private memo and note-taking tool alongside Codex's existing conversation compaction mechanism. Dot takes important notes and wakes up at appropriate times to conduct background research on relevant topics, gradually learning your working details and collaboration relationships; private notes and conversation compaction are distinct tools, and OpenAI does not directly train models on proactive research and private notes; if information from them enters eligible conversations or tasks, your data settings still apply.

2. You Keep Talking While It Dispatches Different Tasks to the Right Environment

What many developers care most about is: how do the underlying environments actually divide labor between Dot and Codex?

In reality, Dot itself is not confined to a single device. While maintaining an ongoing conversation with the user, a Dot can launch sub-Agents and independent tasks in parallel and feed the results back afterward; users can open the corresponding Codex thread at any time to collaborate with Codex directly, or leave Dot to coordinate it. For instance, when brainstorming an idea during a phone call on the go, you can have Dot directly launch a Codex Cloud task.

These three computing environments each serve distinct roles:

Dot's Cloud Computer (Dot's PC)

Dot possesses an independent virtual machine and browser in the cloud. Cloud tasks can continue running even after the user's computer is turned off, and the machine can persist state between uses. In addition to retrieving information, it comes preinstalled with engineering and multimedia software including Blender, FreeCAD, GIMP, Godot, and Kdenlive. The cloud browser operates in an isolated session and does not inherit the logged-in state of the user's local browser.

Your Local Computer (Your PC)

After granting authorization via the desktop app to connect your local computer, Dot can access local files, tools, and your authenticated local browser environment (including computer-use capabilities). This requires the local computer to remain connected to the internet and the ChatGPT client to remain running. Connecting your local computer does not mean exposing all conversations across all applications to Dot.

Codex Cloud Environment (Codex Cloud)

For coding tasks that require specific repositories and configurations, Dot can spin up independent tasks within the user's preconfigured Codex Cloud environment. These cloud coding tasks can likewise continue running when local devices are offline.

In real-world use, Dot gradually adapts to the user's habits and preferences. For example, the author personally prefers delegating backend modifications or quick copy tweaks to the cloud, while running major frontend UI changes on his local laptop; the author notes that Dot gradually learns and adapts to these preferences over repeated interactions; this describes his personal workflow.

The author also offered a possible scenario: you are working on a task through Dot in local Codex, and your calendar shows your commute is coming up shortly; Dot might suggest moving the work to the cloud to continue. It considers not just how far the task has progressed, but also what environment can keep working next. This refers to proactively offering suggestions; an actual switch still requires setting up a cloud task, and should not be construed as existing tasks and full execution state migrating over automatically.

3. Connecting New Messages with Past Tasks, Alerting You to What to Do Next

Proactivity is a defining difference brought by Dot. Beyond what the user explicitly delegates, Dot can pull together information from connected apps and tools in the background to proactively track work progress without requiring human prompts at every step. For instance, a new message in Slack or a new decision in meeting minutes might relate to an item discussed last week; Dot can connect the two and bring it to the user's attention.

• Case 1: Tracking a PR deployment The author had Chopper track the full deployment of a PR to production, notifying him once it was confirmed that the fix could be retested. This eliminated the friction of repeatedly checking status manually, delivering the core update the author actually cared about: confirmation that it was time to retest.
• Case 2: Early warning of potential risks at a conference During DevDay, the author was troubleshooting in Slack to see whether issues encountered during the opening keynote would affect the tech demo in his own presentation. Chopper noticed this, while also spotting another ongoing incident that the author had not yet seen. Chopper cross-referenced the incident details, presentation script, and schedule to alert the author to check the corresponding portions of his demo.

• Case 3: Automated feedback investigation and fix validation An engineer on the Dots team granted their Dot even broader permissions: not only monitoring feedback channels, but also proactively investigating reported issues, spinning up Codex fix tasks, checking CI (builds and tests automatically run after code commits), and addressing failures. When major issues arise, Dot flags them for the engineer to step in; because these tasks are themselves Codex threads, the engineer can jump in directly and pick up the work at any time.

What Can Be Delegated, and What Must You Decide Yourself?

With this degree of freedom granted to an AI Agent, will it spin out of control, modifying code erratically or sending rogue messages?

The key lies in distinguishing what it is researching from what you authorize it to actually do.

1. Default Guardrails, Proactive Research, and Action Review

Proactive Research conducted by Dot in the background is strictly read-only. The tools used for this research cannot send messages, alter application content, or control your browser or computer. Any follow-up actions that could affect accounts or shared information must undergo authorization and automated review: the system combines user instructions, application permissions, custom rules, and built-in security mechanisms to determine whether an action is permitted immediately, requires your approval, or must be handed off for you to perform personally (such as changing passwords).

Simply connecting Dot to messaging channels like Slack does not automatically set up monitoring; you must explicitly specify the channels to observe, the events to watch, and the ongoing responsibilities to fulfill. For ongoing, long-term follow-up tasks, Dot can autonomously decide when to pause and wake up to continue work without needing a fixed schedule configured for every single follow-up check; however, recurring work executed at fixed intervals still requires an explicitly saved schedule.

2. Defining Which Actions Require Asking First via Rules

Custom rules provide optional controls for specific long-term behavioral requirements, allowing you to choose among four behavior modes for specific actions in settings:

• Take action without asking: Executes the designated action directly without requesting approval.
• Take action when you say so: Executes directly when you explicitly request that action; otherwise asks for confirmation immediately before acting.
• Ask before taking action: Must request approval prior to performing the designated action (such as sending a message to a customer).
• Hand off to you: Requires handing the action off to you to complete personally (such as deleting shared project files).

Custom rules are instructions that Dot strives to follow, but oversights remain possible. These rules do not themselves grant access permissions to applications or computers, cannot bypass built-in security protections, and cannot waive mandatory confirmations (such as authorization confirmations when performing a new login using saved passwords).

3. Control, Takeover, and Terminating Work

If you need to take over actions on the cloud PC, you can open that cloud PC and select Take over, directly using your mouse and keyboard; once finished, select Return control to let Dot resume.

When you need to pause or terminate tasks, the scope of different controls varies:

• Pause: Only stops the main task that Dot is currently running; it is not a global emergency stop, so it neither stops all dispatched delegated tasks nor cancels future scheduled runs.
• Activity (Activity panel): To stop independent subtasks dispatched by Dot, you must open the respective task in the Activity panel to inspect and terminate it individually.
• Scheduled (Scheduled panel): Canceling recurring tasks with fixed cycles requires navigating to the Scheduled panel to disable or delete them individually.
• Separation of calls and tasks: Ending a voice call only terminates the voice conversation itself; dispatched background tasks can continue running.

Quota Policies and How to Get Started

Availability and Quotas

• Availability: Currently rolling out gradually to eligible accounts. This includes Pro 100, Pro 200, and Pro 500 (for users aged 18 and older located outside the EEA, the UK, and Switzerland); Business Premium and Enterprise are rolling out globally (Enterprise is disabled by default and requires workspace admin activation).
• Quota consumption rules: Conversations with Dot do not count toward ChatGPT usage limits; work tasks initiated or managed by Dot within Work or Codex count toward the respective product's usage limits as usual. User subscription plans include deep work quotas, with extended limits provided during the first month after launch. What the author describes as "virtually unlimited" primarily refers to everyday conversation with Dot itself not consuming standard quotas, and should not be interpreted as allowing unlimited consumption by background work tasks.
• Entry points and connectivity requirements: Creating a Dot must be done in the desktop app or desktop web interface; creation is not supported on mobile web; once created, the same Dot can be used in updated mobile apps. Connecting a local computer is an independent setting; connecting a local PC is not a prerequisite for interacting with Dot on mobile.

Recommended Incremental Onboarding Path

If you have already built your own follow-up loops or use an overarching thread to coordinate multiple work threads (a thread of threads), you have already solved part of the organization and follow-up challenge. The author also acknowledges that such users might not immediately feel the benefits of Dots in the short term, as the product is still in its early stages.

To get started, consider the "anti-to-do list" concept from Claire Vo cited by the author: identify tasks you repeatedly handle or monitor yourself, and teach them to Dot one by one. The goal is not to have it generate an even longer to-do list for you, but to reduce repetitive work you must personally shoulder once processes, outcomes, and permissions are clearly defined. Here is a practical approach:

1. Pick high-frequency tasks with clear acceptance criteria: Select things you do several times a day or week where you clearly know what a qualified result looks like (such as periodically comparing API changes or following up on specific checks).
2. Demonstrate the workflow and refine it together: Walk Dot through the operational process once, refine and polish the results together, and only then formally hand the task over to it.
3. Provide comprehensive resources and clarify permitted actions: Supply relevant links and necessary access permissions, clearly specifying in conversation the scope it is allowed to handle, or configuring custom rules to define when it must confirm with you.
4. Clarify the collaborative relationship: The author sums up the working dynamic as "delegate to my Dot, collab with Codex". Dot can similarly take on complex tasks without being confined to routine chores; when you need in-depth discussion or paired troubleshooting on specific problems, talk directly with Codex; for ongoing tracking, multi-task dispatch, and coordinated follow-through, let Dot handle it.