I Want Scout to Win. Codex Just Works. Cursor Might Beat Them All.
My working ranking of AI agents for Microsoft 365 collaboration, and the Scout remote we're building to change the outcome.
View all my published articles
Here’s an uncomfortable admission from a Microsoft guy. The agent I want to win is not the agent I rely on. And the one I rely on may not have the highest ceiling. That tension says more about the state of AI agents in 2026 than any benchmark. It’s also why we’re building the remote Scout is missing.
Here’s what actually inspired this piece. I had been living in Codex. Then I finally got Scout running, then hooked it to Grok 4.6, and that’s where it started to sing. Codex is my go-to today, but once our remote is in place, we’re moving fully over to Scout. The winning combination, as I see it, is Grok 4.6 or a newer model, combined with Scout, Work IQ, and native plugins, which are now an industry standard.
Add Cursor, and you have the three agents that currently define how I think about AI in Microsoft 365 collaboration. They’re excellent for very different reasons, and this ranking doesn’t use intelligence as its criterion. It uses the only benchmark that has ever mattered in my thirty years of shipping technology into real organizations. Which agent becomes part of the way you work?
One disclosure before we start. We have skin in this game. My team is building the remote-control surface Scout doesn’t have, and we’re publishing that work in the coming weeks. Read the Scout section knowing we’re not neutral about the outcome.
Microsoft Scout: The One I Want to Win
Scout’s positioning is the strongest of the three. Introduced at Build 2026 as the first of Microsoft’s always-on “Autopilot” agents, it’s a desktop AI that acts across files, shell, browser, and dev tools. Through Work IQ, it also reaches Teams, Outlook, OneDrive, and SharePoint, the Microsoft 365 data that defines my customers’ days. It sits closest to where collaboration work actually lives. For those of us inside the Microsoft graph all day, that isn’t a feature. It’s home-field advantage.
The dealbreaker is simple. Scout has no remote. Microsoft’s own FAQ says it plainly. Scout is a desktop application for Windows and macOS because it needs the local file, shell, and browser control that phones can’t provide. Teams on mobile is the interim surface, and it’s a workaround, not a workflow. When I step away from the machine, I can’t hand Scout work. In 2026, an agent I can only delegate to from a desk has one hand tied behind its back. (And getting to the desk isn’t free, either. Scout is a Frontier preview today, sitting behind an Intune policy, an admin attestation, and both a Microsoft 365 Copilot and a GitHub Copilot license.)
So rather than wait, we’re building our own remote for Scout, a way to hand it work, watch its progress, and approve its actions from wherever I happen to be. We’ll publish what we’re building soon. Weeks, not months. With a remote in place, Scout becomes a very different proposition.
Codex: It Just Works
It just works.
That line has survived every draft of this piece, because it’s the whole review. Codex is the agent I rely on right now, and the reason is the shape of the workflow, not the spec sheet. Give it a job, let it work, come back, inspect, adjust, keep moving. Delegation instead of babysitting.
The real hand-off for me is Dynamics 365. I give Codex a list of people, and it runs them through an enrichment pipeline we built in Azure on Claude Fable 5, then lands the results in our CRM and our line-of-business systems. We connect Codex to that pipeline through a skill, and we’ll do the same with the other agents.
Codex doesn’t need to win every feature comparison, and it doesn’t. It wins the loop. Assign, walk away, return, review. The entire workflow holds together end to end, and that matters more than any benchmark score. The agents that change how you work are the ones you stop thinking about.
Cursor: Maybe the Most Powerful
Cursor may have the highest ceiling of the three. When I push agents across a full project (multi-file changes, refactors, tests, the whole board at once), it’s the most flexible instrument I’ve used. Capability has never been Cursor’s problem.
The story around that capability changed while I was writing. SpaceX announced its $60 billion all-stock acquisition of Anysphere, Cursor’s parent, on June 16, and completed it on August 14, the day I finished this piece. Cursor is now a wholly owned subsidiary of the newly public SpaceX.
A side note on Grok Bot. SpaceXAI, the division formerly known as xAI, opened the Grok Bot beta on August 11, and in the workflows I’ve pushed through it, it’s holding its own with Cursor’s agents. We connected Grok Bot to Microsoft 365 and ran real Dynamics workflows through it. Impressive is the honest word. Notice the wiring, though. Access runs through SuperGrok Heavy, Cursor Ultra, or Cursor Teams Premium. That’s Cursor’s subscription ladder, not a standalone xAI plan. With the Anysphere deal closed, one family now owns the most powerful IDE agent and a general agent that already works my Microsoft stack. One more detail stung. Grok Bot shipped with an iOS app on day one, and Android is on the way. A brand-new agent launched with exactly the surface Scout is missing.
The honest caveat is that potential isn’t the only thing that matters. Cursor’s ceiling is the highest here. Whether it becomes part of my daily loop is a different test, and it hasn’t passed that one yet.
Honorable Mention: The Coworks
One entry, two products. Anthropic’s Claude Cowork and Microsoft’s Copilot Cowork share a line because they share an engine. When Microsoft shipped Wave 3 of Microsoft 365 Copilot in March, it said so itself, describing a product built in close collaboration with Anthropic that brings the technology behind Claude Cowork into Copilot. Same engine, two wrappers.
The wrappers matter. Claude Cowork is desktop-first, with reach into local files and the browser. Since July, its sessions also follow you to the web and the Claude mobile app, running in the cloud while your laptop stays shut. Copilot Cowork runs cloud-side inside your Microsoft 365 tenant, with Work IQ context, Enterprise Data Protection, and admin controls; Gartner notes it gives up local computer use in the trade. Tenant-first shops will happily make that trade. And fair is fair. With mobile in place, the honorable mention already runs a version of the remote loop I’m chasing for Scout.
So why aren’t the Coworks in the top three? Because neither has earned a slot in my daily loop yet, and that’s the only ranking criterion this article uses. This list is a snapshot, not a verdict.
The Pattern Worth Naming
Step back and the ranking reveals something bigger. Microsoft increasingly ships the governed layer on someone else’s engine. Scout runs on the open-source OpenClaw runtime, with Microsoft contributing policy conformance back upstream. Scout’s own release notes even list Claude Sonnet 5 and Claude Opus 4.8 among its models. Copilot Cowork runs on the technology behind Claude Cowork. And on my own desk, Scout runs Grok 4.6, a Microsoft agent on a Musk model, quietly handing the pattern its third engine. In each case, Microsoft’s contribution is the part enterprises actually buy: Entra identity, compliance, auditability, admin control. The runtime and the model are commoditizing; the governed wrapper is the product. I’m not the only one reading it this way, but if you’re wondering where the moat moved, that’s where.
The Real Competition Is the Agent Workflow
So reframe the question. Not “which model is smartest.” They’re all smart, and they leapfrog each other monthly. The question is which agent can actually become part of the way you work. Scout has the natural advantage of where it lives. Cursor may have the highest ceiling, and its new owners are consolidating fast. Codex is the one I rely on, because the delegate-and-return workflow simply works.
Today’s Ranking
Scout is the one I want.
Cursor may be the most powerful.
Codex is the one that just works.
Once the remote works the way I want, we move to Scout full time, and this competition gets a lot more interesting. Everyone else has shipped the loop, with Grok Bot on day one and the Coworks in July. Scout is the holdout, and the opportunity. Follow along. The Scout Remote publication lands in the coming weeks.
The Agents in This Piece
Microsoft Scout (Frontier preview, get started on Microsoft Learn)
Grok Bot (beta announcement from SpaceXAI)
Claude Cowork (getting started guide)
Copilot Cowork (now in Microsoft’s Frontier program)
Paul J. Swider is CEO and Chief AI Officer at RealActivity, a Microsoft Partner specializing in mission-critical AI for healthcare systems. He has 30+ years in healthcare technology, has trained over 3,000 engineers across GE, IDX, and Microsoft, and is the founder of BOSHUG, the Boston Healthcare Cloud & AI Community spanning 50+ countries.