Today Microsoft pauses Copilot Tuning. Agents keep working. Unfinished jobs get tossed. If you've got one running, that's today. If you don't, enjoy not being able to hide a bad permission behind "we're still tuning."

Microsoft 365 Copilot gets a new Microsoft 365 admin center control Sep 25 for advanced agent and plugin uploads: everyone, specific Entra users/groups, or nobody. Admins have until Oct 25 to choose. After that, advanced-package uploads are enabled by default unless restricted. Basic uploads and first-party or third-party built agents are out of scope. If unrestricted advanced uploads are not your governance model, set the boundary before the default sets it for you.

Microsoft 365 Copilot session sharing is on by default. Full session or single response, org-scoped link, snapshot at share time. It does not grant access to referenced M365 files. The control is tenant-wide only. No Entra group scoping. Very enterprise. One switch for everyone.

Self-serve Jira and Confluence connectors hit GA in September. Users connect with their own credentials, and accessible content starts surfacing in Copilot Chat and Search. They are enabled by default. Admins get a review window. Miss it, and your connector strategy starts after the project wiki is already searchable.

Copilot Studio is moving credential oversharing from warning to enforcement. Agents and flows that still ride maker or system credentials can be blocked at design, publish, and share. Waiting for publish to fail is not a migration plan. Inventory connections and give every connection a real identity before governance becomes the bouncer.

Copilot Studio is adding a hard pause before an agent runs a specific tool: human approval. Per-tool, per-agent. Approve, approve for the session, or deny. Inline in Teams and Microsoft 365 Copilot. Prompt instructions say please behave. This says ask first.

Microsoft wants *.cloud.microsoft and copilot.cloud.microsoft open for reliable Copilot. Partial allow lists inside that wildcard are unsupported. If you blocked Copilot to keep personal Microsoft accounts out, use Entra Tenant Restrictions instead of a URL block. Confirm the redirect timing in Message Center or with Microsoft. Do not invent your own deadline.

Sep 1: GitHub Copilot harness agents and workflows created before Aug 3 started burning Copilot Credits for real. Dev and trial environments too. M365 Copilot licenses do not cover this harness. Authoring, preview, evals, and runtime all count. If you have not set an agent-level monthly limit with stop-on-cap in the Power Platform admin center, this week is when the bill shows up.

Agent 365 can finally export who actually used agents in the last 30 days. UPN, agents used, sessions, last activity. From Agents > All Agents > Export. Inventory without adoption is cosplay. This is the first report that looks like a licensing conversation.

Agent Academy’s Learn MCP lab is the useful version of grounding an agent in docs. You point Copilot Studio at learn.microsoft.com/api/mcp. Three tools: microsoft_docs_search, microsoft_code_sample_search, microsoft_docs_fetch. The instruction that matters is the one that says use those tools, and if they miss, send the person to Learn. Don’t invent a cake recipe. Indexing a SharePoint dump is not the same as a live library card.

Microsoft says Copilot Studio skills are the Agent Skills open format, same SKILL.md you’d write for Claude Code or GitHub Copilot. Portable on paper. I have not seen a clean import without a rewrite.

Starting July 2026, Copilot Studio creates an Entra Agent ID for every new agent. You cannot opt out at the environment level. Classic agents keep their old identity. The product finally treated identity as default. The audit trail still has to prove which one wrote the Dataverse row.

GitHub Copilot harness went GA in Copilot Studio on Aug 3. You pick it at create time. You cannot move that agent onto the standard harness later. New agents start burning Copilot Credits the same day. Classic topic agents keep running. The runtime got a lot more ambitious. The identity model still has to catch up.

Where are you building agents this month? Copilot Studio, Agent Builder, or a flow in a trench coat?

You can’t govern an agent you can’t see. That’s the whole dashboard.

We spent 20 years saying apps are not users. Agents: cool, I need both, a mailbox, and not System Admin. Entra Agent ID is the first product that believes that sentence.

Copilot Studio talking to Dataverse over MCP isn't abstract. It's the signed-in user. Same rights. Same tables. When that connector hiccuped on Aug 20, practitioners got a flat "database connection" error and the agent went dark for everyone on it. Fabric Data Agents are the other lane. You ask something like "revenue by region last quarter," it turns into DAX or SQL against a governed semantic model, under Entra. Docs are blunt: no create, no update, no delete. The annoying failure mode there is quieter. Truncated rows. Wrong source if you wire too many. I care more about whose credentials and whether it can write than about which product name is on the agent.

Microsoft's Kernel Memory is the RAG pattern everyone keeps rebuilding by hand. Ingest files → chunk → embed → ask in natural language → get citations back to the source. The part worth stealing into your Copilot Studio / Azure AI Search agents: > answers come with RelevantSources, not vibes > tags can filter by user/collection so RAG respects tenancy > the pipeline is the product: extract, partition, vectorize, store Also worth saying out loud: the repo is archived research now. Learn the pattern. Don't treat it like a supported Microsoft SKU.

User | App | Agent. Four questions under each: who is it, what can it write, who sees the log, who kills it. Same quiz. New column.

Agent needs Dataverse. Do you give it a user, an app, or Dave's old service account? "We'll figure it out" now has a product name.