OpenAI Dots Can Manage Your Coding Agents. Keep the Merge Button.
Maestro Brief · Published by Maestro Mojo
2026-10-01
Maestro’s take. OpenAI’s new dot looks less like a chatbot and more like the project manager sitting above your coding agents. Give it responsibility. Do not give it sole authority to declare its own work safe.
OpenAI introduced dots at DevDay 2026. A dot is an always-on GPT-6 Astra agent with its own cloud computer and browser. It can keep working between conversations, use connected apps, run background agents, and bring decisions back to you.
For developers, the important part is not the cute face. According to OpenAI’s tasks and memory documentation, a dot can create Work or Codex tasks, continue existing local Codex tasks, and follow up on work it delegated. That puts a persistent coordination layer above individual coding sessions.
TL;DR
- A dot can manage ongoing work instead of waiting for a fresh prompt.
- It can start parallel background agents and create or continue Codex tasks.
- It can work in the cloud while your computer is off. Local work still needs your connected computer online with ChatGPT open.
- A completed task is not proof that the job was done correctly.
- Pausing the dot does not automatically stop its delegated tasks or cancel its schedules.
- Conversations with the dot do not count toward ChatGPT limits, but Work and Codex tasks it starts use those products’ allowances.
What changed
A normal coding-agent session has one job: fix this bug, review this branch, or build this feature.
A dot can own the responsibility around those jobs. It can notice a new bug report, start a Codex task, check the result, send a follow-up, and ask you when the next step needs judgment.
That is useful because software work is mostly handoffs. The first patch is rarely the end.
It is also where risk grows. The coordinator can remember decisions, touch several systems, and keep work moving after you leave. A vague instruction now has a longer life.
A practical developer setup
Do not say:
Keep our app healthy and fix whatever breaks.
Say:
Check the connected bug queue each weekday. For new reproducible bugs, open a Codex task in the prepared cloud environment. Ask it to add a failing test, make the smallest fix, run the named test suite, and prepare a draft pull request. Send me the test result, changed files, and remaining uncertainty. Do not merge, deploy, close the bug, or message the reporter without my approval. Alert me immediately if credentials, production data, billing, or security are involved.
That prompt gives the dot five things it needs:
- A source of truth.
- A narrow trigger.
- A definition of done.
- Evidence to return.
- Actions that remain human decisions.
Do this
- Give the dot one ongoing responsibility, not a vague department.
- Name the repository, environment, queue, and documents it should trust.
- Require evidence: tests, changed files, screenshots, logs, or a draft PR.
- Define what deserves an interruption and where to send it.
- Keep merge, deploy, deletion, spending, and external messages behind explicit approval until the workflow earns more trust.
- Review Activity and Scheduled separately. They are separate control surfaces.
Do not do this
- Do not let “task completed” mean “result verified.” OpenAI explicitly says completion alone does not confirm the requested result was achieved or delivered.
- Do not assume every delegated agent sees the dot’s whole history. New tasks receive selected instructions and context, not every conversation.
- Do not assume Pause stops everything. OpenAI’s controls guide says delegated tasks and recurring schedules must be stopped separately.
- Do not treat memory as a perfect record. A dot keeps notes, not a complete transcript.
- Do not assume changing ChatGPT saved-memory settings changes notes the dot already made. OpenAI documents saved memory and dot notes as separate systems.
Why Maestro users care
Coding agents are getting good at individual jobs. Maestro’s inference is that coordination is becoming a larger share of the real cost: starting the right work, preserving context, checking results, and deciding what happens next.
Dots move that coordination into a persistent agent. That can reduce waiting and repeated prompting. It can also turn one fuzzy request into a chain of fuzzy actions.
The simple rule is: let the dot coordinate the work, but require another control to prove the work. For code, that means tests, review, and protected merge rules outside the agent’s own report.
Who can use it
Dots are rolling out gradually, so eligible users may not see them immediately.
OpenAI currently lists access for Pro 100, Pro 200, and Pro 500 users over 18 outside the European Economic Area, United Kingdom, and Switzerland; Business Premium worldwide; and Enterprise worldwide with administrator enablement. Enterprise dots are off by default.
The dot’s conversation itself does not count toward ChatGPT usage limits. Deeper Work and Codex tasks it creates or manages consume those products’ allowances as usual.
One thing to try
When dots reach your account, assign one read-mostly workflow first: triage new bug reports and prepare a daily list of reproducible issues.
Do not let it change production. Measure whether its evidence saves you time for one week. Then add one bounded action, such as preparing a draft PR.
The bottom line
The useful part of dots is not that they run forever. It is that they can remember a responsibility and keep the handoffs moving.
Use them as coordinators. Keep proof and consequential approvals outside the coordinator’s say-so.
Sources considered
Published October 1, 2026 · Tags: OpenAI, ChatGPT, AI Agents, Codex, Agent Workflows