Interview with Chirag Shah (University of Washington) on AI agents

Are AI agents just another buzzword — or are they poised to transform the way companies operate? Mecalux spoke with Chirag Shah, Computer Science Professor at the University of Washington and a leading expert in artificial intelligence, to explore how AI agents are reshaping everyday life, business operations, and the future of human–AI collaboration.

• Not everyone understands what an AI agent is …

  AI agents aren’t new. They’ve been around as long as AI itself. At their core, an agent is any entity— software or hardware — that can gather inputs and take actions independently. In the history of AI, the earliest vision of agents was rooted in robotics: machines operating in the real world, making choices without waiting for human instructions.

  Today, when we talk about AI agents, we mostly mean software agents. But the principle is the same: they make decisions, carry out delegated tasks, and act on our behalf. A thermostat is an agent — it makes choices on your behalf, like turning on the heat or AC when a threshold is crossed. At the other extreme, a self-driving car is also an agent, just a much more complex one.

• What’s driving the rise of AI agents today?

  Agents are having a resurgence because they’re now powered by foundation models. Unlike the earlier generation — think Siri or Alexa, limited to narrow tasks — today’s agents can draw on large language models, vision systems, or multimodal tools.

  These agents don’t stop where their own capabilities end. If they can’t do a calculation, they reach out to a calculator; if they don’t know stock prices, they query a trading platform or search the web. Some can even write code to solve problems. On top of that, foundation models are getting better at reasoning and connecting dots. That’s why agents today are already far more capable, effective, and practical than the previous generations.

• So, they can handle more complex tasks?

  Exactly. Their ability to use tools, reason, and even work together with other agents through frameworks like AutoGen or LangChain means they can now tackle challenges that were once out of reach. We’re seeing multi-agent systems already solving tasks that used to be too complex for one system alone.

  That said, with every wave of agents, there’s been a sense of “this is it; we’ve figured it out” — and then we hit walls. Even today’s agents face those walls. That’s why I wrote Agents are not enough: to highlight that building more capable systems isn’t the only challenge ahead.

• In your paper, you introduce the ideas of sims and private assistants.

  

  We know from previous generations of AI agents that if all they do is provide the weather, play music, or turn off the lights, they don’t generate enough value for customers to keep using them, let alone pay for them. Real value emerges when agents can handle complex tasks. But even then, complexity alone isn’t enough. End users also expect personalization and proactivity. Imagine asking an agent to book a trip between Barcelona and Toulouse. A truly proactive, personalized agent wouldn’t just suggest flights; it might also recommend trains, anticipating your preferences.

  But achieving personalization raises privacy concerns. An agent may need access to sensitive information such as medical records, finances, or personal routines. If we simply throw this data into the hands of public agents, we risk breaches of trust.

  My proposal is a dual system with two types of agents: public and private. The private assistant belongs entirely to the user, with no hidden agenda. It taps into what I call “sims”: representations of different parts of your life, from work and finances to health and personal routines. Using this information, the private agent can design highly personalized tasks while keeping your data under your control. Only when you approve does it turn to public agents to get the job done. This way, the assistant doesn’t act randomly and doesn’t need to keep asking endless clarifying questions because it already knows you enough. Instead, it delivers proactive, tailored recommendations, and the risk of making a mistake is much lower than with a public agent.

• What about software agents versus embodied (robotic) agents?

  When I was a student, the definition of an agent was based on embodied systems — robots out in the world, using sensors and signals from the environment to make decisions. But over the last decade, the focus has shifted to software agents. Now, as robotics has advanced, some of those earlier ideas are becoming practical, especially in controlled settings like factories, where robots are deployed as agents.

  For most users, though, the agents they encounter every day are still software ones: shopping assistants, productivity tools, or customer service bots. Over time, these lines will blur. Ideally, people won’t think about whether a task is being done by a robot or a piece of software. They’ll simply say, “I want this task done,” and the right kind of agent — embodied or digital — will handle it.

  What I see as most promising is a future of human–AI collaboration

• How do you see AI agents changing work and human–machine collaboration?

  I don’t believe we’re heading for a future where agents simply replace humans. Some people argue for that, but I don’t think it’s realistic or ideal. At the same time, I also don’t think it makes sense to resist agents altogether. What I see as most promising is a future of human–AI collaboration.

  Take medicine as an example. When a case is complex, you often have a team of doctors with different specialties working together to make a diagnosis. Tomorrow, that team could include not only human doctors but also AI agents tuned to specific domains. In that setup, the agents would bring speed, scale, and analysis, while humans contribute judgment, empathy, and accountability. Together, they could diagnose cases much faster and more accurately.

  This approach emphasizes augmentation. Agents shouldn’t strip humans of control, especially in high-stakes situations like healthcare or transportation, where mistakes can be extremely costly. Instead, they can empower people to achieve more, better, and faster.

  

• Many businesses are already experimenting with AI agents in operations, logistics, and supply chains. How realistic is it to let agents make decisions that affect revenue?

  Agents excel at speed and scale. They can process huge volumes of data and make decisions far faster than humans. But performance isn’t only about quantity. Think of a factory assembly line: if you’re mass producing, you need automation to scale and achieve cost efficiency. Yet humans remain essential, not only to design the system and apply best practices but also to ensure quality, control, and accountability.

  While agents may accelerate processes, accountability still rests with the company. You can’t tell regulators or customers, “The agent did it.” In the end, the business owner is responsible for the outcome, whether the decision was made by a human or an agent.

• Researchers are testing AI agents in supply chains to negotiate and reach consensus. How close are we to seeing AI agents negotiating on behalf of businesses?

  It’s already happening in several domains. In the advertising industry, for instance, these negotiations take place all the time and in fractions of a second. There are so many parameters to consider for ad selling and placement that human-to-human negotiation is just not feasible. We didn’t call them agents at the time, but that’s essentially what they were: processes making decisions in real time for both sides, clients and service providers. The same is true in stock trading. Agents are executing orders at speeds humans can’t match.

  We’re now moving toward a future where every business could have its own AI agent negotiating on its behalf, interacting directly with agents across the table. This is not science fiction; it’s already happening. More businesses can now take advantage of these capabilities to speed up negotiations. Agents can already handle negotiations that require countless parameters in real time, something nearly impossible for humans. This is one domain where this technology can play a very important role.

• Some logistics projects are combining digital twins with AI agents to simulate entire supply chains. Is this already happening in other industries?

  

  Many systems that we now describe as agents or digital twins have existed for years in industries like finance, healthcare, and retail. In retrospect, we can look back and say: yes, those were examples of agent-to-agent negotiation or multi-agent systems, even if they weren’t labeled that way. People solving real problems didn’t always call them agents. They just built processes that worked.

  Finance is a good case in point. A huge share of transactions already happens autonomously through agents, often rule-based systems that are explainable, auditable, and scalable — crucial for a heavily regulated industry. Healthcare offers another lesson: AI has been integrated as a tool, but not as a replacement. Doctors might rely on AI for analysis or note-taking, but the ultimate responsibility still rests with humans. And in retail, dynamic pricing is a well-established example. Platforms like Uber and DoorDash constantly adjust prices in real time based on demand and availability. Humans oversee the system, but they don’t drive the calculations.

  So, supply chains can also benefit from these approaches. The real opportunity now is to learn from the successes and failures of other industries.

• What advice would you give to business leaders considering AI agents?

  Even though much of my research and consultancy focuses on AI agents, I often begin with a caution: “Are you sure you want agents?” In my experience, companies that dive into AI with the wrong motivation — such as “everyone else is doing it, so we should too” — sooner or later pay the price. I’ve seen many cases where projects don’t work out or become too costly.

  At the end of the day, what really matters is running the business, solving problems, delivering value to customers, and complying with regulations. My advice is simple: focus on the problem you want to solve. Start with the solution; don’t worry about the name. Stay focused on the actual business and look for real solutions that address your needs. In the long run, that will serve you far better than being swept up by the latest trend or hype.