Copilot Studio Agents Now Create Files Users Can Download. Plan for Where They Go.

Microsoft updated the Learn article Files the agent creates (GitHub Copilot) (preview) on 28 October 2026. It describes a capability that is easy to read as a convenience: an agent built on the harness can now produce a Word document, a spreadsheet, a slide deck, a PDF, a chart, or a code file in the middle of a conversation and hand it back to the user with a Download button. The governance point is simple. Every harness agent now has a file output channel that the maker does not switch on, that works in Teams, Microsoft Copilot and web chat, and that the documentation does not yet tie to labels, DLP, audit or any admin setting. Treat it as a new way for data to leave a conversation, and plan for it before the preview becomes the default.

The article is still marked as prerelease documentation, and the page title carries the “(preview)” label. It also still uses the GitHub Copilot harness name, which Power Platform Engineer covered when it became the Microsoft Copilot harness.

The capability at a glance

A file output channel that nobody turns on

Most Copilot Studio capabilities that move data are visible in the agent definition. A maker adds a knowledge source, a connector, a tool or a workflow, and an admin can find it in the solution, the inventory or the data policy. Created files work differently. The article says the agent produces a file “when the agent needs to write a file to complete a turn”, and that makers do not configure the behavior. There is no file creation toggle in the article, no environment setting and no Power Platform admin center reference.

That matters because a file is a different shape of output from a chat answer. A response in Teams is read and scrolled past. A spreadsheet built from an agent’s knowledge sources can be downloaded, forwarded, uploaded somewhere else and kept indefinitely. An agent that is grounded in sensitive SharePoint content or Dataverse tables can now package what a user is allowed to see into a tidy, portable document on request. The access boundary does not change, but the effort required to extract data at volume drops sharply.

Labels and DLP: the documented gap

The companion article, Attach files to a conversation (updated 12 October 2026), states that “User-attached files support Microsoft Purview labels.” The created files article makes no equivalent statement. It does not say whether a file the agent writes inherits the sensitivity label of the source content, carries any label at all, or is visible to Purview DLP. Admins should not assume that the label protection documented for inputs also covers outputs.

Power Platform data policies govern connectors, not the files a harness agent writes with its built-in packages. If your control strategy for agents leans on data policies, the comparison in Power Platform Data Policies vs. Microsoft Purview DLP is worth revisiting with created files in mind. Until Microsoft documents label behavior for created files, test it in a non-production environment with labelled source content and record what you observe.

Retention: 28 days, measured from the last activity

Created files are “stored with the conversation for the same period as user-attached files”, which both articles give as 28 days after the last activity in the conversation. After that, files are deleted from the conversation. Two points follow. First, “last activity” is not defined, so a conversation that a user keeps returning to can hold files for longer than 28 days in practice. Second, the deletion applies to the conversation copy only. Once a user has selected Download, the file lives wherever they saved it, under whatever controls apply there.

Records and compliance teams should decide whether agent-generated documents count as business records, and if so, where users are expected to save them. The article does not describe export, recovery after deletion, or eDiscovery for created files.

Cost and preview status

The article does not give a credit rate for file creation. It repeats the general note that the harness uses usage-based billing and that agent use might consume Copilot Credits. Generating a multi-sheet workbook with charts is likely to involve more work per turn than a short answer, so file-heavy agents deserve a line in your cost model. The approach in The Harness Is the Price Tag applies here: measure real conversations in a test environment before you forecast.

Preview status is the other cost. Preview features can change without the notice period that generally available features get, and the article says so directly. Production agents that depend on created files should have a fallback, such as a workflow that writes the document to a governed SharePoint library.

What the documentation does not cover yet (as of 9 November 2026)

• Any admin control, environment setting or tenant setting to turn file creation off, or to limit it to some agents.
• Whether created files carry Microsoft Purview sensitivity labels, or are visible to Purview DLP. Label support is documented only for user-attached files.
• Audit events for file creation or download, and whether Agent 365 or Microsoft Purview audit records them.
• Where the files are physically stored, and how data residency applies. The article says only that they are “stored with the conversation.”
• Who else can see a created file in a shared channel, such as a Teams group chat, and how access is enforced.
• What counts as “last activity” for the 28 day retention window.
• The Copilot Credit cost of creating a file, and a general availability date.
• Whether a maker can disable the built-in packages or built-in skills. The skills overview (updated 28 October 2026) says built-in skills “are provided by the harness, rather than uploaded or authored by a maker” and does not mention a way to remove them.

Admin checklist

1. Inventory agents built on the Microsoft Copilot harness, and flag those grounded in confidential knowledge sources or Dataverse tables.
2. In a test environment, ask a harness agent to export labelled source content to Word and Excel. Record whether the downloaded file carries a label and whether DLP reacts.
3. Check your audit tooling for any event when a file is created or downloaded. If none appears, note it as a monitoring gap.
4. Agree with records management whether agent-generated documents are records, and tell users where to save them.
5. Add agent instructions for sensitive agents that limit what the agent should export, and test them with evaluations.
6. Include file-heavy scenarios in Copilot Credit testing and environment allocations.
7. Treat the feature as preview in production designs, and keep a governed fallback path.
8. Watch the article and the Agent 365 control plane for admin controls, and revisit this checklist when they arrive.

Sources

• Files the agent creates (GitHub Copilot) (preview), Microsoft Learn, last updated 28 October 2026
• Attach files to a conversation (GitHub Copilot), Microsoft Learn, last updated 12 October 2026
• Skills overview for agents (GitHub Copilot), Microsoft Learn, last updated 28 October 2026
• What’s new in Copilot Studio, Microsoft Learn, last updated 4 November 2026