Agentic (Premium)

With Copilot, Microsoft promises a “side-by-side” experience in which AI improves the work you’re doing in traditional apps and experiences. But wave two of the AI revolution adds not just automation but new agentic capabilities in which AI services–called agents–work independently on your behalf and then only bother you when they’re done. If you’re a developer, you may think of this shift as the “async of AI,” where you no longer have to wait on AI to come back with results. It will catch up with you when it can.

Here’s how Microsoft describes it.

“Agents … can now operate independently, dynamically planning and learning from processes, adapting to changing conditions, and making decisions without the need for constant human intervention. These autonomous agents can be triggered by data changes, events, and other background tasks—and not just through chat!”

I love the exclamation mark and the implicit implication there that “triggering” AI through a chat interface was somehow an ideal way to get work done. It recalls the early days of text-based adventure games in which the player faced two challenges: Winning the game, of course, but also figuring out which terms the limited game parser would accept. AI, like those games did, is improving, and rapidly, and just as text adventures evolved into graphical adventure games like King’s Quest, with ever more impressive graphics and parsers, AI, and the ways in which we interact with it, are likewise changing.

To me, these new agentic capabilities are likewise reminiscent of Microsoft’s initial push for .NET almost 25 years ago. As originally conceived, .NET was a way to make the Internet proprietary, more powerful, and more centralized. The key consumer services Microsoft envisioned at the time were .NET My Services, originally code-named Hailstorm, though only one major service, Passport, made it from vision to final product. (That it still survives today as the ubiquitous Microsoft account is rather incredible, given how central it is to the Microsoft consumer experience.) Hailstorm was a lot of things, but I think of it as the connective tissue of a proprietary Internet that never materialized. But the closest it ever came to what we’re seeing today was a component called .NET Alerts, which was designed as a system for managing and customizing alerts.

It’s possible that .NET My Services was ahead of its time. It’s also possible that it was simply too proprietary and Microsoft-focused, at a time when Microsoft was being torn down by antitrust regulators and the world was becoming more heterogeneous and open. Whatever your take, it didn’t happen, and as we shift forward 20+ years, there are similar uncertainties around AI and the impact it will have on our lives. But Microsoft, for now, is all-in. And in its AI/Copilot-centric view, agents are the next step.

I appreciate Microsoft admitting that agents aren’t new: It has offered developer and IT tools for creating agents and agent-like services for years. But large language models (LLMs) make agents inherently more powerful and useful, Microsoft says, and the two together are uniquely capable.

We’ll see how that pans out, but there’s little doubt that agents can be useful: Consider a simple but useful online service like Google Flights, which can ping you when a certain flight you’ve been tracking reaches a specific price threshold. That’s undeniably useful. But also undeniably simple, a basic “if-then” computations working on a finite set of data. And Microsoft’s vision for AI-powered agents is far more sweeping.

This opens the door to these new agents running into the same complexity that doom “If-this-then-that”-type services, as they can take multistep assignments that involve all kinds of criteria. Where a real personal assistant is general, a Jack of all trades that does it all, AI agents are specific. Each does one thing. And if this system works, we’ll all have many agents out there, working on our behalf.

The high-level promise will sound familiar, since it’s the same as with any AI functionality: Agents will “free” you to do the work that matters most to you. To the cynical, this is like claiming that spelling and grammar checking had a similar impact on your life, and fair enough. But AI capabilities scale far beyond that, or can, and we have useful and interesting generative AI capabilities today that speak to that. With agents, the AI “partners” with you on daily tasks, but not interactively, and not in real-time.

To achieve this, these new agents require memory, entitlements, and tools, Microsoft says.

Memory provides continuity. Each time you request something from an agent, it happens in the context of your previous requests, and not in isolation. Entitlement is a fancy word for permissions: Like Cortana and other previous generation digital personal assistants, these agents can’t do anything for you unless you let them. And the tools bit refers to the apps and other software the agents need to take actions on your behalf.

Interesting.

What Microsoft is describing there sounds a bit like the orchestration capabilities that I’ve long felt have been necessary but missing from AI since the current boom times began. Indeed, if you go back to how Stevie Batiche described how AI capabilities would be introduced into the ecosystem at Build 2023 about 18 months ago, you’ll find that he literally described “outside applications,” the third of the three AI app structures he discussed, in terms of agents and orchestration.

“AI will orchestrate across multiple apps, plugins, and services, functioning more as an agent,” he said at the time. “If you take a step back, the Windows shell itself is an orchestrator. In fact, may be one of the most powerful orchestrators across apps, across content, across the [Microsoft] Graph. Imagine with AI and natural language, you start to see glimpses of the opportunity. And it is here when you get intelligence that is functioning not just at granular details, but at the higher levels where you get a mixing of both tactics and strategy, you get both vision and execution. It’s like a Copilot of Copilots, a very powerful application structure.”

A Copilot of Copilots. That’s also interesting. And maybe even terrifying, especially to those who approach the AI capabilities that Microsoft has added to familiar tools warily. Copilot is one thing, it appears on the side when summoned, and you use it, or you don’t. But now Microsoft is bringing agents to Microsoft 365. And these things won’t just appear when summoned. They’ll appear when they’ve accomplished something on your behalf. And that could mean you’ll use it whether you want to or not.

Microsoft provides a few examples of this will work. A real-time speech-to-speech translation service in meetings called Interpreter. An HR and IT help desk service called Employee Self-Service that will resolve issues and determine whether you’re effectively using your employee benefits. Dynamics 365 agents with sales, supply chain, finance, and customer service functions. Agents trained on the content in your organization’s SharePoint site that can find what you need in seconds. Developer agents built with LLMs and small language models (SLMs), and a new developer SDK. A recruiting agent for LinkedIn that will help with hiring.

Microsoft is sensitive to the safety concerns, as expected, but it also acknowledges fears that increased AI use will result in job losses. Many of the agents it’s developing include “human in the loop” approval schemes, and because we’re suddenly living in the future that Isaac Asimov envisioned in his Robot series of stories, authors can review records to see which actions an agent took and why. One wonders about the formal creation of a tribunal process for punishing these agents when they inevitably run amok.

OK, I’m kidding about that, but the backlash against manually summoned generative AI tools like Copilot was so strong and sudden that it’s inevitable this advance will be met with a similar pushback. And perhaps rightfully so. Agents can solve problems, reason to some degree, and automate tasks without our ongoing involvement. Microsoft describes them as a “paradigm shift in terms of how work gets done,” but it’s difficult not to imagine that fewer humans in the chain are key to that efficiency.

So what does this automation look like?

In Microsoft 365, agents will “unlock” capabilities in the Microsoft 365 apps and services via new Copilot Actions that work like “set and forget” fill-in-the-blank prompts you fill out once and then never think about again. Whether this more Mad Libs or Choose Your Own Adventure is currently unclear. But Microsoft cites examples like daily action item summaries at the end of each workday, input outreaches to team members for a newsletter you send out each week, and automatic summaries of previous interactions with participants of a coming meeting. You will be able to delegate–yes, really–tasks to agents, freeing you to focus on other work instead of the rote, time-consuming stuff.

Put simply, agents–or what we might call agentic capabilities–are appearing everywhere in the Microsoft stack. And in those few places that Microsoft overlooks, or with special complex cases that none could foresee, we’ll be able to create our own agents, whether we’re developers or not, spin them up, and let them loose on the world, God help us all. This is, in Microsoft’s worldview, a good thing, of course. But again, it’s perhaps too easy to see the dystopian possibilities here as well. Too easy to imagine an AI agent telling me, sorry, Paul, I’m afraid I can’t do that, once it figures out that I’m the weak link in this chain of inefficiency.

As I write this, I have nothing but questions. I suspect I will get some answers to these questions as this week’s Ignite Conference unfolds. I certainly have plans to scan through the available session replays so I can learn more. But I suspect that when this time next year comes, there’ll be an agent that will do that for me, an agent that will understand what I care about the most and proactively queue up the most relevant session videos.

Whether that scares or excites you says a lot about how accepting you are of our new AI overlords. For me, it’s a bit of both. But maybe I’m coming down with something. I’m feeling a little … agentic at the moment.