The AI Agents Are Here — What's Left for Humans to Do?

I have a friend who's a university professor researching AI. He recently supervised a graduate student who, within the bounds of academic ethics, leaned heavily on AI agents to help with the research. The thesis came together smoothly — and then the student suddenly blurted out: "Wait, so what was the point of me coming to grad school?" That sense of deflation is probably a lot like what plenty of engineers felt the first time they used AI to help write code and watched it knock out an entire feature in one shot. My friend, as the advisor, was genuinely at a loss for words for a good while.

The AI Agent Era Arrives — Human Value Gets Redefined

What makes people feel like they themselves have become useless — this is fundamentally different from the past, and the biggest reason is the rise of the AI agent. About a year ago, when I'd bring up AI agents with clients, most people would just stare blankly, not knowing what I meant. Now, their employees have more or less all used one, whether they realize it or not — because the moment you use any of the major AI labs' products, you naturally end up using an agent.

For ordinary people, AI used to mean going to talk to a chat box, getting your answer, and still having to go do the actual work yourself. The AI agent, by contrast, is closer to the kind of robot we've all imagined since childhood — it can finish the task itself, check its own work, deliver it on its own, and even mutter to itself and generate its own next instruction. So the relationship between people and AI has become genuinely subtle. When a machine can run a process start to finish, where exactly is the value of the person standing next to it?

But for an AI agent to actually function, it turns out to need help too — just like a person, it can't operate all alone in the world with no tools at hand. It needs a whole stack of underlying infrastructure.

The good news is that this year, those underlying protocols have basically converged, and development is starting to have real standards: the Model Context Protocol (MCP) handles agent-to-tool connections, A2A handles agent-to-agent communication, and the Agentic Commerce Protocol (ACP) and Universal Commerce Protocol (UCP) govern commercial transactions. These three aren't fighting each other — they're stacking as separate layers.

Of the three, Anthropic's MCP has essentially already won the tools layer — adopted across the board by Anthropic, OpenAI, Google, and Microsoft, effectively locking down the position of connecting agents to outside tools. On a recent trip to the US, I saw that Google's heavily promoted A2A is already leading the agent-to-agent collaboration layer, with more than 50 technology partners on board, and shopping-oriented protocols like ACP and UCP are starting to see adoption from e-commerce vendors too. The infrastructure has taken initial shape — meaning machines can now talk to other machines and delegate tasks to each other on their own. The agent has become a genuinely new software layer, and every major player, you'll notice, already has one in hand.

Beyond infrastructure, the clearest sign of this growth is simply the amount of money pouring in. The agent category itself is exploding — the market was reportedly worth about $7.8 billion in 2025, and is projected to hit $52.6 billion by 2030, a compound annual growth rate of roughly 41%. Gartner, meanwhile, predicts that by 2028, a third of enterprise software will have agentic AI built in, and roughly 15% of our everyday work decisions will be made autonomously by agents. In other words, the pieces of work that used to require human effort are going to get taken over by agents, one chunk at a time.

From General to Specialized: Vertical Agents Become the Next Wave of Winners

So where's the real opportunity in the AI agent business? Silicon Valley has had a consistent view here: the real opportunity lies in "vertical" agents — finance-specific, accounting-specific, legal-specific, and the like. The reason is that these fields are packed with deep, tacit domain knowledge and heavily dependent on proprietary enterprise data, which a generic model can't simply swallow on the fly. As for the general-purpose, do-everything-a-little agent, in the end that's probably only a game the capital- and compute-rich giants can afford to play.

The evidence backs this up: the fastest-growing companies over the past two years have almost all been vertical players deeply focused on a single industry — Harvey in legal, Ambience in healthcare, Sierra in customer service (which hit $150 million in annualized recurring revenue in just eight quarters, reportedly the fastest in enterprise software history). Their moat, when you strip it down, is proprietary data plus deep enough integration that outsiders can't copy it quickly. It's no surprise that top-tier accelerators and VCs are now betting substantial money specifically on these vertical agents, wagering that the next Salesforce-caliber winner will grow out of one of them.

So whose AI agent should you use? Taiwanese business owners love to say "we'll just build it ourselves" — but the license fee they save usually gets paid back double, because the tuition of building an in-house team is far more expensive than it looks on paper.

An MIT study last year found that AI systems built by outside vendors succeed at twice the rate of systems built in-house through trial and error. And the same MIT report found that 95% of enterprise AI still hasn't generated a return. So enterprises need to be patient, because the execution loop is the most expensive part of the process — and if that work loop can be handed to an AI agent, success there can produce a genuinely massive shift in a company's cost structure.

As AI agents keep getting more capable, I suspect more and more people, on some late night after a long day at the office, will stare at the screen watching an agent that finished the job itself and even double-checked its own work, and feel a sudden jolt of "who am I? Where am I? What am I even doing here?" But what the agent takes over is execution — what it leaves for humans is the question. That graduate student, honestly, was already holding the best answer in his hands: the person capable of asking "so what am I even here for" is, by that very fact, still very much needed.