# Dots vs Codex: Who Coordinates, Where the Work Runs
**Plutonous** | October 5, 2026 | 6 min read
> A Dot can coordinate Codex, but a cloud browser, a cloud coding environment and a connected local machine are different workplaces. Follow the task to the deliverable.
Tags: OpenAI, ChatGPT Dots, Codex, AI Agents, Cloud Development, Developer Tools, Task Orchestration, Git Worktrees
---
**TL;DR: Treat Dots and Codex as 2 cooperating roles: coordination and execution. OpenAI documents that a Dot can delegate to Codex, while connecting a personal computer does not transfer its cloud browser session.[1][2] Choose the workplace explicitly, then verify the changed files or finished artifact rather than accepting task completion as delivery.**
The most expensive misunderstanding in an agent workflow can be a simple one: “on my computer” means different things to different participants. A user sends a request from a phone. A coordinator organizes the work in the cloud. A coding task changes files on a connected host. The conversation is mobile; the repository is not.
OpenAI's current documentation makes Dots and Codex complementary. Dots can maintain responsibilities and delegate; Codex supplies coding workflows with a concrete project environment.[1][4] Our October 5 analysis is that the useful comparison concerns task placement and accountability. Asking which product is smarter skips the operational decision that determines whether work can happen at all.
### Why This Matters Now
An agent can coordinate a task without owning its execution environment. Before delegating, name the repository or source files, the machine or cloud environment, the expected output and the evidence required for acceptance. Those choices make a broad request reviewable.
*Cover: AI-generated editorial artwork about dispatching work between environments. It is not an OpenAI interface, deployment diagram or measurement of product performance.*
## The Coordinator: Give the Dot a Responsibility
The Dots getting-started guide describes delegating parts of a request to ChatGPT Work or Codex and inspecting that work in Activity.[1] That suggests a useful division of labor: a Dot can own the continuing concern, while a specific coding task owns a bounded change.
Consider an illustrative checkout bug. Ask the Dot to track the issue, gather the relevant reports and coordinate a proposed fix. Give the coding task the repository, starting state, reproduction steps and acceptance criteria. The point of coordination is to keep those inputs connected to the outcome, including new information that arrives after implementation starts.
The strategic value lies in reducing the owner's attention cost. A coordinator that notices an unresolved dependency can be more useful than another worker writing code. But that benefit depends on a disciplined handoff: the child task needs enough context to make the right change without reconstructing the entire business conversation.
## Cloud Work: A Browser and a Coding Environment Differ
A Dot's cloud computer has its own files and browser sessions.[2] Codex Cloud, meanwhile, runs coding tasks from a configured environment, with separate workspaces for tasks and review of changed files and test results.[4] “Cloud” identifies location; it does not specify capabilities.
The current Codex Cloud environment guide explicitly lists computer and browser use as unsupported. It also says repository skills are available, while personal local skills are not synced.[5] A Dot's browser capability therefore cannot be assumed to exist inside its delegated coding environment.
Our practical conclusion: split the job at the resource boundary. Research in a suitable browser can produce a source brief. A coding environment can consume that brief and change a repository. If the required step involves a desktop application, choose an environment that actually exposes it. Delegation should move instructions and evidence deliberately, rather than relying on the word cloud to hide the transition.
## Local Work: Preserve the Repository's Starting State
OpenAI documents local environment setup scripts for new worktrees and reusable project actions such as running tests.[6] Its worktree guide says the repository and commands remain on the computer or remote development environment that holds the project.[7] A new checkout is a place to work, not proof that dependencies or uncommitted resources are ready.
For our hypothetical bug, local execution is useful when reproduction depends on the developer's installed tools or files. A worktree can separate the proposed change from foreground development. Ask the worker to identify its checkout and starting revision so the reviewer knows which version was tested. The valuable output is a change against a known baseline, not a floating collection of edited files.
A sensible task brief also states what must be preserved. Existing edits, generated assets and project-specific setup may matter to the result. Clear instructions here save the coordination layer from managing an avoidable recovery operation later.
## Remote Control: Your Phone Is the Control Surface
The Remote guide says remote access uses the connected host's files, credentials, permissions and tools, and supports reviewing diffs and test results.[8] Worktree documentation likewise says worktrees do not run on the phone.[7] Mobile access changes how the owner directs work; it need not change where execution happens.
Dots documentation makes another distinction: trying a connected computer after a cloud-browser problem may create a separate local task, without transferring the browser session.[2] Treat that as a change of workplace. A login or open page on one machine does not establish what another task can see.
This matters for planning. Keep a machine available when a job depends on it, and request a status that names the executing host. If the local step is blocked, a useful coordinator can still prepare source material or clarify requirements. It should report the dependency instead of declaring that remote access somehow removed it.
## Task Ownership: Context Must Survive the Handoff
OpenAI says delegated tasks have their own conversations, receive selected instructions and context, and remain associated with their original computer or cloud environment. Changing a Dot's selected computer does not move an existing task.[3] Ownership follows the task, not the latest location selected in the coordinator.
Our recommendation is to request a small execution record: task identity, working location, intended deliverable and current blocker. Use that record when sending follow-up instructions. “Keep working” is ambiguous if several tasks exist; “continue the checkout fix in its existing repository and rerun the regression check” is actionable.
Control follows the same logic. OpenAI's Dots controls distinguish pausing the main task, stopping delegated work and canceling a schedule.[9] When priorities change, inspect the actual work items. A quiet coordinator is not evidence that every worker has stopped.
## Acceptance: Follow the Work to the Artifact
The desktop app documentation emphasizes inspecting real files, while the Dots task guide warns that a completed run does not establish delivery.[10][3] That is the accountability test for this entire arrangement.
For the checkout example, ask for the diff, relevant test result and a reproducible demonstration. For a report, ask for the saved document and source notes. For an analysis, ask for the input window, calculation and reviewable table. These are editorial examples of acceptance evidence, not claims that the product automatically produces each item.
### Completion Is a Claim to Verify
Our recommendation: identify where the result was written, open it, and check it against the request. A coordinator's summary, a worker's finished status and a usable deliverable are separate pieces of evidence.
The real story isn't Dots replacing Codex. It is a coordination layer making execution easier to direct. The advantage goes to teams that know where work runs and insist on evidence where it ends.
## Sources
1. [Get started with your Dot](https://learn.chatgpt.com/docs/dots/getting-started)
2. [Connect computers and apps](https://learn.chatgpt.com/docs/dots/computers-and-apps)
3. [Dots tasks and memory](https://learn.chatgpt.com/docs/dots/tasks-and-memory)
4. [Codex Cloud](https://learn.chatgpt.com/docs/cloud)
5. [Cloud environments](https://learn.chatgpt.com/docs/environments/cloud-environments)
6. [Local environments](https://learn.chatgpt.com/docs/environments/local-environment)
7. [Git worktrees](https://learn.chatgpt.com/docs/environments/git-worktrees)
8. [Remote connections](https://learn.chatgpt.com/docs/remote-connections)
9. [Control your Dot](https://learn.chatgpt.com/docs/dots/controls)
10. [ChatGPT desktop app](https://learn.chatgpt.com/docs/app)
*Last updated: October 5, 2026*
---
*Source: [LLM Rumors](https://www.llmrumors.com/news/dots-vs-codex-cloud-local-task-execution)*