Living docs in ChatGPT Pages: accuracy rules for agent edits
This research library uses AI-assisted source research and drafting. Linked sources support product claims; analysis and proposed exercises are our interpretation. Unless an article documents a test and its results, do not read it as a hands-on review or an independently verified benchmark.
ChatGPT Pages is a new document type where people and agents write, research and edit together. OpenAI launched it at DevDay on 29 September 2026 for Pro, Business and Enterprise, as part of ChatGPT Space. A page can carry instructions for an agent: what to update, which sources to use and when to check.
That makes living documentation possible. It also means a status page, a handbook or a decision log can change while nobody is reading it. This guide sets out the rules we would agree before letting agents maintain team docs.
Which documents should agents maintain in ChatGPT Pages?
Not every page should be agent-maintained. Sort documents by the cost of a wrong line.
| Document type | What an agent may do | Human owner | Review cadence |
|---|---|---|---|
| Weekly status or release notes | Draft and update from linked sources | Team lead | Each run, until trusted |
| Meeting notes and action items | Summarize and propose actions | Meeting organizer | Same day |
| Team handbook | Propose edits only | Named maintainer | Monthly |
| Decision log | Suggest wording, never change a decision | Decision owner | Every edit |
| Customer or security facts | Read only | Data owner | Not agent-maintained |
The pattern is simple. Agents may summarize evidence. People own decisions and commitments.
Accuracy rules for agent-maintained pages
Write these rules into each page's agent instructions block, and into the team's shared guidance.
Every factual line needs a source link. The Space getting started guide models this: its example prompt asks ChatGPT to preserve facts and source links and to list open questions at the end.
Mark proposed items as proposed. The agents docs walk through a launch brief where a date differs from the linked plan. The fix is to ask the agent to describe the date as proposed and awaiting approval, while keeping the date and source unchanged. That is a good default for any unconfirmed claim.
Agents propose, owners accept. Ask for change shows a proposed edit for review before you accept it. Show attribution displays who contributed, when available. Use both on decision pages.
Say when a source is missing. An agent that cannot reach a connected tool should state that on the page, not fill the gap with a guess.
Agree who edits which section. The docs note that collaborators share page content but may use different agents, connected tools or instructions. Two editors can produce two versions of the truth.
Scheduling is a separate control
The launch material says pages can stay updated from connected tools. The docs are more precise. Keep Updated is not available at launch, and writing a cadence in a page does not schedule anything. Recurring updates run as scheduled tasks or, for shared work, as Team Tasks.
Team Tasks bring their own governance points:
- Enterprise admins grant two permissions: Create teams, and Create and manage team automations.
- Team Tasks run in the cloud with the team's service account and configured connections, not the creator's personal access.
- Connected accounts may reach information a member's own account cannot.
- Tasks can write to connected systems if the tool and connection allow it, subject to action approval.
- New team members see all earlier runs and generated files. Runs cannot have separate access restrictions.
- Pausing or deleting a task should not be relied on to stop an active run.
- Team Tasks use workspace credits, with team spending limits separate from user limits.
Approval points before agents update your docs
Run this checklist before the first unattended update:
- Name the page owner and write the accuracy rules into its agent instructions.
- Confirm which account each connection uses and what it can read and write.
- Test the task prompt manually in a regular chat, as the scheduled tasks docs advise.
- Review the first few runs against their sources before anyone relies on the page.
- Decide who can pause the task and who reviews the Activity view.
Admins can review team activity through the Compliance API without joining a team. Confirm which records exist before you rely on them for an audit.
What should you measure?
Count corrections per page per week, and track how many came from people spotting an agent error. Note how long a wrong line stayed on the page before someone fixed it. Check whether readers still ask the owner for the same information.
A falling correction count with steady readership suggests the page is trusted. A page nobody opens does not need an agent. Our measurement guide explains how to run that comparison fairly.
The same review discipline applies to agent-written code. Our reviewable change checklist covers that side, and the free methodology guide explains how we split delegation and ownership.
Next step: pick one status page, name its owner, and add a source-link rule to its agent instructions today.