MIT News
Q: What is agent-based AI and how does it differ from generative AI models such as ChatGPT and Claude?
AND: Agentic AI is artificial intelligence that takes action in the world. These actions can be physical, such as manipulating a robot, or digital, such as booking a flight. On the other hand, we think of generative AI as inventing stories, poems, works of art and images, rather than taking actions for us.
The word “agent” is just a brand name. This usually means artificial intelligence that will aid people interact with an app, website or the physical world. Most of the agents we encounter today are digital agents, such as customer service agents, who you can talk to about product complaints.
Most agent companies operate the same few AI models that give them the ability to take action and remember what happened. The agent starts with a basic, generative AI system like Claude. Companies then place different packaging around this base model of their product or application. These wrappers can be specific tools that an agent can operate, and these tools are application-specific. Perhaps the agent has access to a calculator so he can solve math problems, or maybe he has access to a more complicated demanding drive and operating system so he can remember company financials and past business negotiations.
The biggest challenge in developing agentic AI is the lack of training data. If I want to create a system that can go online and book a flight for me, it seems pretty elementary. But we don’t have much data that explains exactly how to do it – where to move the mouse, which buttons to click, what to do if something goes wrong, or how to call someone and negotiate the price of an airline ticket. One way to train such a system is to ask an AI agent to visit airline websites, try different solutions and see what works and what doesn’t. These environments are complex to model, so the agent often has to learn through trial and error.
Q: What are the promising applications of agentic AI?
AND: I think the area where we had the most success was coding agents. This is something that evolved from generative artificial intelligence. People trained language models on code and then could predict what a human would do to solve a coding problem. Additionally, the agent can learn this by going through a feedback loop where it tries different solutions and checks whether it has given the correct answer. As long as the AI agent can check the answer, it can conduct a trial-and-error loop until it finds a good strategy.
However, there is always a balance between automating decision-making and simply helping and informing people. Analytical artificial intelligence methods, such as systems that aid predict possible decision outcomes, are not agentic in nature, but provide a wealth of information for human decision-makers. In cases that are high or safety-critical, such as medical, security, high-level business policy, etc., the technology may not be ready for AI to completely automate these processes, or we may not even be comfortable with it.
Q: Are there risks we should be aware of when using AI agents?
AND: One massive area of risk comes from the fact that it is often very effortless to hire agents to do certain types of work for you. With coding agents, you can “vibe code” and simply have an agent create the code for you so you don’t have to do the demanding work yourself. There’s a massive risk that because it’s so elementary, people won’t put enough effort into making sure it’s doing it right. Errors will appear, private data will be leaked – this is already happening.
Agents are not perfect in the sense that they can make mistakes because they are not well trained and do not know what to do. But even if they are very competent, if a human does not operate them appropriately or gives instructions that are too vague, the AI agent may make a mistake because the human made a mistake. I think if people are less committed to thinking about all the consequences, we will be more susceptible to making these mistakes.
An additional aspect is the risk of losing qualifications. It’s unclear how far this will go, but as we rely on agents to do homework, program, and perform math, we may lose the ability to do it ourselves, and we may lose that ability too soon because the technology is not yet ready to fully automate these processes.
Q: What does the future hold for agentic artificial intelligence?
But then again, maybe an exceptionally good coding model could act as a puppeteer for communicating with sensors, actuators, and web APIs? Perhaps once you have a superintelligent reasoning system that understands math, language, and code, you can give it a camera and a keyboard and it will figure out what to do in the spatial domain. Will the next wave of AI just be Claude with sensors, actuators and tools, or will it be something built from the ground up in a novel way? This is the most essential question that many people working in the field of artificial intelligence are struggling with today.