Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Teams

Microsoft 365 Copilot connectors can now write

Written November 13, 2026

Connector write actions are GA in Chat and Cowork, with approve-once or remembered approval. Researcher's connectors stay read-only.

What changed

Copilot connectors spent their first life reading: they brought information from other applications into a conversation so Copilot could answer with it. On November 9, Microsoft announced that connector write actions are generally available across Copilot Chat and Cowork. Where a connector provides the tools, a person can now ask Copilot in plain language to create, update, or delete information in the connected application, without leaving the conversation. Microsoft says write actions are coming to Copilot in the Office apps as well, and gives no date for that step.

The mechanics matter more than the headline. Copilot prepares the action and shows what it intends to send. Nothing reaches the connected service until the person approves it. Once approved, Copilot carries out the change with that person's own permissions in the other system, so it can only change what the person could already change themselves.

Approval is per tool

By default, Copilot asks for permission before it creates, updates, or deletes anything. A confirmation card appears on screen naming the tool and the connector, along with the details being sent, so the person reviews the actual payload rather than a summary of it. For each tool, there are two choices:

The setting is per tool, not per connector, and it can be changed at any time in chat Settings, under Sources, then Tools. Microsoft's defaults split by risk: read tools start at Always allow, while new write tools start at Needs approval. An admin who wants a tighter posture does not have to build it; it is the starting state. What an admin does need to decide is which connectors are enabled at all, because remembered approvals accumulate quietly once people start clicking Always allow on the tools they use every day.

What people are expected to do with it

Microsoft's announcement walks through six partner connectors. Atlassian Rovo creates and updates Jira work items and Confluence pages, so decisions from a meeting transcript become backlog items with acceptance criteria attached. Notion pages and databases can be created and updated from chat, turning a discovery call into a spec with a status set. Canva creates designs and applies team feedback to them. G2 builds campaigns from a request and its supporting brief, with the setup shown for approval before anything launches. iManage drafts a new work order from the governing agreement without changing the original contract. BioRender turns a scientific concept into a figure built from its templates.

The pattern across the examples is consistent: the conversation supplies the context, the connector supplies the destination, and the approval card is the seam between the two.

The admin side

These are federated connectors: they reach the external system live through the Model Context Protocol, and Microsoft's documentation notes that no external data is indexed into Microsoft 365. Admins govern them in the Microsoft 365 admin center, where connectors can be enabled or paused, limited to specific users or groups through staged rollout, and capped by the tenant-wide User access setting. A user sees a connector only when the publisher type is allowed, the connector is enabled, and they sit inside both the user access population and the connector's rollout, with a Microsoft 365 Copilot licence. In the admin center, admins can open each MCP-based connector to see which of its tools can create, update, or delete data, which is the review to run before enabling one.

The boundary that did not move

One surface stays read-only. The Researcher agent does not support write, update, or delete actions, and federated connectors remain read-only in that experience even now that Chat and Cowork can write. A team that runs deep research and operational work in the same tenant gets two different risk profiles under one product name, and guidance to staff should say so plainly.

Capabilities also vary by connector, and by configuration and permissions inside it. GA describes the platform capability, not a promise that every connector in the gallery writes today.

Related playbooks

Sources

Back to: Teams playbooks hub · Also: Copilot joins Slack and Teams for real · Webflow's MCP server inside Microsoft Copilot