The Agentic Age
LLMs, Agents, and the Future of AI
No, not that sort of AI agent
A few months ago, I wrote two articles on AI, here and here. They focused on the long arc of AI development, its relationship with robotics, and the global challenges AI is well-suited to address. This time, I want to talk about a somewhat more technical aspect of AI, the rise of so-called AI agents. These are specially developed AI “entities” that are designed to accomplish tasks in the real world. AI as we know it essentially consists of something called “Large Language Models” (LLMs). These LLMs train on a massive database of human-produced content (everything from the Bible to the Walmart website) to gain an understanding of the relationship between words, sentences, images, and eventually ideas and concepts. They create an illusion of understanding by expertly pulling together words or pixels in the order that seems right to any given question, based on every question and answer that’s ever been fed into it by developers. They “think.” AI agents are powered by LLMs, but aim to accomplish much more. Agents use their LLMs to go out into the world and do tasks you assign to them. In short, it is the difference between a thinking machine and an action machine. To understand this dichotomy, let’s zoom out.
Once upon a time, when dinosaurs roamed the earth, if you had a question you wanted answered, you might go to a library. You would use the dewey decimal system to find the right book, then you would use the table of contents to find the right chapter, and skim the headings until you found what you wanted. This is manual direction.
Then, Google and other search engines came into being. Now, if you had a question you wanted answered, you asked the search engine, and it would send you the best sources available, to help you answer it. This is a direction machine.
Now, we have ChatGPT and other AI systems. If you ask AI a question, instead of directing you to the best place to look, it synthesizes every answer given to similar questions, to produce the best answer. This is a thinking machine.
Soon, AI will be given specific tasks. Instead of asking it to answer something, you will ask it to do something in the real world, outside of the website or app you access it from. This is an action machine.
At present, AI is basically a generalist product. There are some models that are better at math than others, and there are some reskinned AIs designed for corporate use that limit themselves to certain fields. Typically however, AI will give answers based on the entirety of the vast back of knowledge it has access to. Agentic AI approaches things somewhat differently. Although its actions are still informed by a vast database of information, an AI agent is a specialist designed to accomplish a specific task.
So, what tasks can an AI agent accomplish? Some are not such a big jump from what most of us use AI for today. With a voice command to your AI agent, you could add a new meeting into your Google Calendar. Then there are slightly larger tasks, like asking an agent to find you a dinner table at one of your 5 favorite restaurants, between 6:00 and 7:30. There are also things that you can’t afford the agent to mess up, like booking, and paying for, all your travel plans for a week-long trip to Europe. Ultimately though, these are all relatively modest personal applications.
Agents could also play a central role in the workplace. In fact, we have already seen some adoption by specific companies. Salesforce recently fired 4,000 workers, mostly in customer service, because agentic AI could do most of the work itself. Soon we will see more advanced agents resolving issues instead of just pointing clients in the right direction, or explaining to them how they can fix the problem themselves. Investors are starting to assign AI agents to independently pursue investment ideas and send finished reports to Portfolio Managers. Once success is proven in these roles, agents are sure to appear more often. There will be autonomously operating cybersecurity agents, code-writing and patch distributing agents, human-supervising agents, and more. Agents will manage the office and order supplies without being prompted, suggest design improvements to streamline aerodynamics, or anticipate changes in demand and alter production. Eventually, agents will play a critical role in science, autonomously devising and executing (perhaps with the assistance of robots) experiments in everything from chemistry to microbiology. Those uses are not going to happen next month, but will be possible in the coming years. When agents of that power are implemented widely, it really will be an AI’s world.
The capacity to run effective agentic AI is not so far away. In fact, ChatGPT is already beginning the rollout of semi-agentic AIs that can be granted access to things like your Google Drive, but cannot modify the documents within. Google’s own AI, Gemini, is rolling out an agentic program called Antigravity, but for now it’s aimed at enabling corporate tasks like software production. These capacities will grow in the coming months, but I suspect that adoption of more advanced features will be uneven. In fact, it seems that the biggest challenges to this new chapter of AI adoption will be non-technical. So far, most of our AI usage has been with LLMs that are in a sense, “trapped.” They can bring information in, or, be fed it, but they cannot “get out.” They can click through a website to read the different pages, but they cannot yet book a plane flight. Slowly but steadily, this is changing in low-stakes areas. ChatGPT is already testing agents that will book you a restaurant table, but that’s about it.
As I see it, the primary challenges to agentic AI are legal and regulatory. Who is liable if any AI mistakenly books 30 plane tickets, when it was supposed to book 3? Can you sue the AI company for the agent’s mistake? No AI company will want to take on legal liability for the actions of agents given access to credit cards and bank accounts. (I doubt banks or credit card companies will be eager to give AIs access in the first place.) At the same time, can you really expect people to want to accept the risk of their AI assistants making all sorts of mistakes? What about fraud, can we be sure that gullible AIs don’t give out personal information while just trying to be helpful? Will companies doing sales online even allow AI agents to interact with them? Or will they ban them from their purchasing software, worried about potential legal exposure if the AI makes mistakes. Given time, agents will become more competent, and a regulatory/legal framework will develop to handle these challenges. Overcoming them could take years, potentially holding back rollout in the meantime.
All of this uncertainty is downstream of a core reality of AI. It can be 95% effective at these tasks, and that won’t be good enough. Even 99% may not be good enough. Looking at the resistance that is building to Waymos in many cities, it is clear that people are fundamentally suspicious of entrusting their lives or finances to autonomous technology. AI companies will make their decisions with two opposing forces in mind. First, AI companies want to roll agents out as quickly as possible. They could be packaged and sold as new products to existing AI users, while deepening AI’s penetration into corporate and personal use markets. The AI companies are desperate to increase use, because they are bleeding horrifying amounts of cash every month. Sam Altman and Co. have shown no signs of slowing down in their multi-billion dollar investment announcement, but pressure from investors to deliver profits grows every month. Every day they wait on an agentic rollout, it will hurt their bottom line. The countervailing force is the fight for public opinion. AI companies are desperately worried about their reputation, and have tried to steady nerves by saying that this is a decade long process, and mass layoffs are neither imminent nor inevitable. (This is probably not true, but regardless.) A chaotic rollout of agents would do a great deal of damage to these companies, both reputational and financial. A successful rollout might be toxic in its own right; if the proliferation of AI agents are directly linked to mass layoffs, at which point a major political movement against AI is likely.
I hope I’ve made it clear that agents are an emerging technology. We cannot yet predict exactly what will happen, or how it will happen. It does seem obvious however, that once agents overcome challenges to adoption, they will change the world forever.