ChatGPT Space and Dots put agents in your documents. Who reviews what they write into your records?
OpenAI's Space and Dots make agents real teammates in shared pages. Accept/undo and auto-review cover prose and actions. The facts other agents read as true need something else.
On 29 September 2026, at DevDay, OpenAI made agents full teammates in the document. People, ChatGPT, Codex and a new kind of always-on agent called Dots now work in the same pages, at the same time, in a product called ChatGPT Space.
That is a real step, and it deserves to be taken seriously on its own terms. It also leaves a question on the table that is easy to miss in launch week. A document is where agents talk to you. A record is what the next agent, the Monday report or the customer-facing app reads as fact. Space answers the first very well. This post is about the second.
What launched
- Space is the team's home in ChatGPT. OpenAI's help center calls it "a home for your pages, files, and shared work with ChatGPT." It holds Pages (editable documents), spreadsheets, presentations, Sites (websites and tools built with ChatGPT), images and uploaded files, grouped into spaces by topic or team. For Pro, Business and Enterprise users it replaces the existing Library. Slides, per TechCrunch, will roll out "in the coming weeks."
- Dots are always-on agent coworkers. VentureBeat describes them as persistent agents "designed to keep working after an employee closes the chat window", each with its own cloud computer and browser. You @mention a Dot or ChatGPT in a page to take the next step. Pages can also carry standing instructions; Sam Altman's example was checking a team channel and updating the page with what it finds.
- Everyone edits the same page, live. "Multiple employees with edit access can work on the same Page simultaneously, with each person using their own ChatGPT against that common document." Where available, "Show attribution" shows who contributed what.
- Generated text is reviewed in place. From the Pages docs: "Review the generated text in place. You can accept it, ask for a change, or undo it."
- Consequential actions get an automatic check. An "auto-review system checks potentially consequential actions against the user's instructions, Custom Rules and OpenAI's built-in safety requirements" before deciding whether work proceeds on its own or waits for approval. Custom Rules "can permit particular actions, require approval or prohibit them."
It is available on Pro, Business and Enterprise, on web and desktop first, with mobile later. OpenAI has not disclosed pricing for additional Dots or extra capacity.
What OpenAI got right
Most "AI in your docs" features so far have been a button: select text, get a rewrite, paste it back. Space goes further than that. It makes the agent a participant in shared work instead of a tool one person holds.
Three design choices stand out:
- Shared context instead of private chats. Each person brings their own ChatGPT, but everyone works against the same page. That removes the most common failure of team AI use, where five people get five different answers from five separate conversations and nobody knows which one made it into the plan.
- Review sits where the text sits. Accept, ask for a change, or undo, in place. You judge generated prose in its context, not in a side panel. That is the right shape for writing.
- Autonomy with a check in front of it. The auto-review step, combined with Custom Rules that can permit, gate or prohibit actions, is a genuine pre-execution control. Always-on agents with their own browser need exactly that, and OpenAI shipped it on day one rather than as a follow-up.
For a team whose output is mostly prose (plans, briefs, meeting notes, launch docs), Space is a strong default, and many teams will not need anything more.
Documents are where agents talk. Records are what agents read.
Now follow one piece of work past the page.
Imagine a Dot with a standing instruction to watch the sales channel and keep the account plan current. On Tuesday it writes that the customer's renewal date moved to March and the contract value is $84,000. The account owner sees the paragraph, it reads fine, and she accepts it.
That paragraph is now doing two jobs. For the humans in the page it is prose: context, easy to reread, easy to correct in conversation. For everything downstream it is a claim. The forecast agent reads the renewal date. The finance report reads the contract value. A Codex job that builds the renewal dashboard reads both. None of them sees the paragraph the way the account owner did, and none of them knows it was accepted in four seconds between meetings.
The two review mechanisms Space describes are real, but they are aimed at different things:
| Mechanism | What it reviews | When | The question it answers |
|---|---|---|---|
| Accept / undo in place | Generated text in a page | After generation, before you keep it | Does this read right? |
| Auto-review + Custom Rules | An action an agent is about to take | Before the action runs | May this agent do this? |
| Record-level review | A changed field value, with its source | Before the value becomes canonical | Is this value true, and who decided? |
The first two are described in OpenAI's documentation and press coverage. As of 30 September 2026, the third is not: we found no description of structured records with a schema, of a field-level diff or proposal-and-merge for agent writes, or of version history for pages. Pages can contain tables, so structure is possible. What is not described is a place where a changed value arrives as a proposal, with its old value, new value and source, before anything reads it as true.
That is not a flaw in a document product. Documents are optimized for being read by people, and a person reading a page brings their own judgment. The gap appears when the next reader is not a person.
Which agents?
The second open question is about who is in the room. Space is built around people, ChatGPT, Codex and Dots. As of 30 September 2026, OpenAI has not described a public API or MCP access that lets agents from other vendors read or write Space content.
That is a reasonable launch scope. It is also different from how many teams work today. A typical engineering or operations team in 2026 runs several agents: Claude Code in the repo, Cursor in the editor, Codex for background jobs, a ChatGPT workspace for research, and a scheduled agent or two on top. If the facts those agents depend on live in one vendor's workspace, the other agents either cannot reach them or reach them through copies.
What to do this quarter
Space will be in many teams' hands within weeks. Before agents start writing into it at volume, four decisions are worth making on purpose:
- Sort outputs into prose and records. A launch brief is prose. A renewal date, a price, a headcount, an approved vendor list, a policy threshold: those are records, even when they first appear in a paragraph. Write the list down; it is shorter than you expect.
- Give records a schema'd home with provenance. Facts that other agents consume should live in fields, with the source that supports them and the agent or person that proposed them. A paragraph can link to the record; it should not be the record.
- Review in proportion to consequence. Do not put a gate on every write. People start approving without reading, and then nothing is really reviewed. Gate the writes that matter: values that change money, commitments, compliance or what a customer sees. Let everything else stay fast.
- Keep the system of record agent-neutral. Use whichever agents do the best work for each job, including Dots. But the place your facts live should accept writes from all of them under the same rules, so switching or adding an agent does not mean moving the truth.
Where Busabase fits
Busabase is an open-source database and workspace for AI agents — agents and people share the same structured data, docs, skills, and apps, and writes that matter become reviewed, trusted records.
It is meant to sit next to a tool like Space, not replace it. Keep writing, planning and discussing in pages. When a piece of work produces a fact other agents will read, send it to a Busabase record. ChatGPT can already do that: through its MCP app it can read Busabase and propose records, and with proposal-level access the result arrives as a Change Request with a field-level diff and its sources, rather than silently becoming canonical (Connect ChatGPT). Claude Code, Cursor, Codex and other agents write to the same records through MCP, Agent Skills or the OpenAPI surface, under the same review rules. Low-stakes writes, and keys that already hold merge rights, go straight through.
It is MIT-licensed and runs locally with one command, no account needed:
npx busabase server
# → http://localhost:15419/dashboard/local
For a side-by-side on scope, sharing, agent access and data residency, see Busabase vs ChatGPT Space. It says where Space is the better choice, too.
Disclosure: I work on Busabase. We are not neutral about where records should live, which is why every claim about ChatGPT Space above is quoted from, or limited to, what OpenAI and the press had published by the verification date. Where something is "not described", that means we looked and did not find it, not that it cannot be done.
Sources: ChatGPT Space help center, Work with pages, ChatGPT Space for Enterprise, VentureBeat on Dots and Space, TechCrunch on the launch. Verified 30 September 2026. Space launched the day before and will change quickly; check the primary sources before relying on any detail.