What it does

Researcher is a reasoning agent that tackles multi-step research tasks. It combines a deep-reasoning model with your work data and, where allowed, the web to produce a structured, cited output — think competitive analysis or a market brief.

Key facts

• It runs an agentic loop: plans, gathers, reasons, and comes back with a longer worked answer rather than a one-shot reply.
• Grounds in your Microsoft Graph content and can pull from the web when policy permits.
• Shipped as part of the Copilot agent line-up, initially via the early-access ("Frontier") programme before broad availability.
• Lives inside Copilot Chat as a pinned agent.
• Federated Copilot connectors (GA per the Sept 2026 release notes): Researcher can pull third-party data live over the Model Context Protocol under the user's own identity. Nothing is indexed or stored in Microsoft 365, and the source system enforces permissions. Researcher is read-only against them; write, update and delete tools aren't available to it.
• Users need a Microsoft 365 Copilot licence and must connect with their own credentials. Admins govern connectors in the Microsoft 365 admin center (Agent and plugin access, User access, and per-connector enablement with staged rollout).
• Oct 2026: a Canva connector lets Researcher include connected Canva content in a task, where the connector is available in your tenant.

When to use / skip

Use it for genuinely multi-source questions where you'd otherwise spend an afternoon reading and synthesising. Skip it for quick factual lookups — that's ordinary Chat, and Researcher's deliberation is overkill.

Configuration decisions

• Whether web grounding is available to the agent, and under what DLP policy.
• Who gets access, especially during any early-access window.

Gotchas

• It's slower by design. Users primed on instant chat replies think it's hung. Tell them it takes minutes.
• Cited doesn't mean correct — check the sources it leans on before you circulate the output.
• Federated connectors only see what the signed-in user can see in the source system, so two people running the same brief can get different reports. If the MCP server restricts by source IP, allow Microsoft 365 Copilot's static egress addresses.
• Output quality depends heavily on how well you framed the task; a vague brief gives a vague report.

Consultant notes

• Feature availability and naming here have moved fast — verify what's actually in the customer's tenant before demoing.
• It genuinely changes the economics of desk research, but only for people who write a proper brief. Coach the prompt.
• Connector sprawl is the new governance question. Decide publisher-type policy in Agent and plugin access before users start connecting accounts.
• Web access is a governance decision, not a default. Settle it before you switch the agent on.

---

Worth revisiting after the next release wave.