AI & Agentic Weekly Update — September 14, 2026

Cowork's September Refresh Puts Cost and Effort in the User's Hands

Microsoft 365 Copilot Cowork got its most substantive interface update since going generally available, and the changes are less about flash than about control. The left navigation now reads Automations instead of "Scheduled," the model picker moved out of the corner and into the prompting box — split cleanly into a GPT section and a Claude section based on what your org enables — and there's a new reasoning-effort slider that lets a user dial a task from "Faster" (a quick, cheaper pass) to "Smarter" (a slower, more thorough pass that pulls more context and burns more credits). For the first time, the person writing the prompt gets to decide how hard the agent works before it starts.

The bigger story is cost visibility. The cost skill now surfaces your individual session cost, the percentage of your monthly quota remaining, the credits you've used, and when the limit resets — and a new View usage button opens a full panel showing your spend against the policy and the entire group's spend. That last part matters more than it looks: on most tenants, users share a spend policy, and if the group hits 100% utilization, nobody on that policy can run Cowork. Consumption is no longer an invisible admin concern; it's now visible to the people who create it.

For organizations, this is the difference between agentic AI as a mystery line item and agentic AI as a managed resource. The effort slider and the usage panel together give you a real lever: you can coach teams to reserve "Smarter" runs for work that warrants it, and users can see the shared tank draining before it runs dry. But visibility only changes behavior if people know how to read it — a slider and a percentage bar mean nothing to someone who was never told what a credit is worth.

What to do: Build a one-page "how Cowork spends" guide for your users now that the panel is live. Explain what the effort slider actually costs, how the shared group quota works (and that one heavy user can lock out the team), and give a simple rule of thumb for when "Smarter" is worth it. Turn the new visibility into a habit before someone discovers the group quota by hitting the wall.

Shareable Skills Turn One Person's Work Into an Org-Wide Asset

The most strategically interesting Cowork change is quiet: you can now share skills with colleagues. Open a skill in the Customize section, click the share dialog, and send it to anyone in your organization so they can add it to their own Cowork experience. Skills — the reusable, packaged instructions that tell Copilot how to do a specific recurring task — were previously something each person built and kept to themselves. Now a well-designed skill can propagate across a team the way a good template or macro always has.

This lands alongside a strict brand adherence toggle in brand-kit skills that forces Copilot to use only approved template layouts rather than inventing its own, and an allow note instructions option that lets Copilot read and respect guardrails you've embedded in slide notes. Custom skills have also fully rolled out inside PowerPoint (via Manage skills, including uploading a skill markdown file) and can now be authored directly in SharePoint, where they save to OneDrive and follow you across sites. The through-line is clear: Microsoft wants skills to be portable, governable, shareable units of expertise.

For a consulting-minded organization, this is the moment skills stop being a personal productivity trick and start being organizational IP. The scarce resource is no longer who can write a skill — it's who knows the process well enough to encode it correctly, and who reviews shared skills before they spread. A sloppy shared skill propagates a bad habit across a department just as efficiently as a good one propagates a best practice.

What to do: Stand up a lightweight review process for shared skills before they proliferate. Pick two or three high-value, repeatable tasks your teams do constantly, build reference skills for them properly, and designate an owner who vets skills before they're shared org-wide. Treat your best skills like published templates — versioned, reviewed, and documented — not like forwarded email attachments.

Copilot Studio Adds Human Approval Gates — Autonomy With a Brake Pedal

Copilot Studio shipped the feature that makes autonomous agents genuinely deployable in a regulated enterprise: human approval gates for agent tool calls. When an agent is about to take a high-stakes action, the gate pauses execution and requires a person to approve before it proceeds — a deterministic guardrail rather than a hopeful instruction. Paired with new credential safety detection, it closes what Microsoft is framing as the enterprise compliance loop: an agent can now run a long, multi-step task on its own while still stopping at the exact points where a human must sign off.

This builds directly on last month's general availability of the GitHub Copilot harness in Copilot Studio — the reasoning-heavy foundation for agents that need multi-tool execution, workflow orchestration, and agent-to-agent interaction. The harness gave agents the horsepower to do serious autonomous work; approval gates give organizations the confidence to let them. Together they answer the objection that has kept most autonomous-agent projects stuck in pilots: "what happens when it does something expensive or irreversible without asking?"

The design lesson here is that autonomy and control are not opposites — the mature pattern is an agent that runs freely across the routine steps and stops deliberately at the consequential ones. Deciding which actions deserve a gate is now a core design skill, not an afterthought. Gate too much and you've rebuilt a manual process with extra steps; gate too little and you've handed real authority to a system that doesn't understand consequences.

What to do: For any agent you're building or piloting in Copilot Studio, write down its "irreversible or expensive" actions — anything that spends money, sends external communication, deletes data, or commits the organization — and put an approval gate on each one before you widen access. Make the gate map an explicit part of every agent's design review, the same way you'd review who can approve a purchase order.

Copilot Becomes a Genuine Multi-Model Platform

A shift that's easy to miss in the feature noise: Microsoft Copilot is now a multi-model platform in practice, not just in principle. The in-box model picker in Cowork now openly separates GPT and Claude families, the latest frontier models from both OpenAI and Anthropic are landing in Cowork and Copilot Studio within days of each other, and Copilot Chat now lets you switch the model on any individual response — regenerate an answer you didn't love and pick a different model without copying your prompt into a fresh conversation. Model choice has moved from a tenant-level configuration to a per-task, sometimes per-response, decision.

This matters because the models are not interchangeable. Different frontier models have real, observable differences in how they reason, write, follow instructions, and handle long context — and the person closest to the task is often best positioned to judge which one fit. By exposing the choice at the point of work, Microsoft is betting that users will learn those differences the way they learned which app to reach for. Copilot Studio makes the same bet for builders: agents can now be pointed at the model best suited to their job rather than whatever the platform defaulted to.

The risk is decision fatigue and inconsistency. A picker with two model families and several options each is powerful for a knowledgeable user and paralyzing for everyone else, and "Auto" only helps until someone needs to know why two colleagues got different answers to the same question. The organizations that get value here will treat model selection as a teachable skill with light guidance, not an unmarked menu.

What to do: Give your users a short, opinionated cheat sheet: which model to default to for drafting versus analysis versus code, when to bother switching, and when to just leave it on Auto. You don't need a policy — you need to lower the cost of a good choice so people aren't guessing every time the picker stares back at them.

Scout Nears GA as Microsoft Rallies Builders at the Agent-a-Thon

Microsoft Scout — the first of the new Autopilot class of always-on agents — is holding in private preview for Frontier organizations with general availability still tracking toward October, and Microsoft plans to offer it as an add-on for M365 E3 and E5 tenants once pricing is confirmed. Unlike a Copilot you invoke, Scout runs continuously in the background across Teams, Outlook, calendar, and your files, operating under its own governed Entra identity. The countdown to broad availability is real, and the governance questions that come with a perpetually-running agent are about to stop being hypothetical for a lot more organizations.

The timing is not an accident. Microsoft is running its Agent-a-Thon as a digital event this week (September 14–17), squarely aimed at getting organizations to build and deploy their first agents — a sign that Microsoft sees the current bottleneck as adoption and hands-on skill, not capability. The tooling has largely arrived; what's missing in most organizations is people who have actually built, scoped, and governed a working agent end to end.

That gap is the opportunity. The distance between reading about autonomous agents and running one responsibly is exactly the distance an organization has to cross this quarter, and the ones who start now — while Scout is still in preview and the stakes of a mistake are low — will be ready when the always-on agents arrive by default. The teams that wait for GA will be learning governance and design under production pressure instead of in a sandbox.

What to do: Use the Agent-a-Thon window as a forcing function. Pick one concrete, low-risk workflow, build a single agent for it this month, and put it through the full cycle — scope, approval gates, model choice, cost, review. The goal isn't the agent; it's building the internal muscle to design and govern agents before Scout and its successors make that skill mandatory rather than optional.

Sources

Back to all news