As big tech developers race to roll out the next generation of artificial intelligence, some people are already enlisting an army of AI helpers to check off their daily to-do lists.
These are called agents: AI systems that automatically come up with a plan and execute it at your request. They can take the form of a shopping assistant, a video gaming coach or even a fully self-driving car.
“This notion that we will have our own personalized assistant, if you will, is one way to think about it,” Tammy Madsen, professor of strategy & innovation at Santa Clara University, said about the rise of these specialized virtual workers.
Last month, OpenAI, the company behind ChatGPT, introduced an AI agent called Operator that it advertised as capable of using the web to complete tasks like filling out forms and ordering groceries for a certain recipe. In a demo of the agent, OpenAI showed Operator booking a dinner reservation, purchasing tickets to a basketball game and ordering pizzas, each time following the user’s specific preferences.
Google on Wednesday also announced the launch of an AI co-scientist, which the company described as a “virtual scientific collaborator” made up of multiple AI agents.
Businesses have been eager to adopt them as well, using AI-powered tools to automate tasks traditionally handled by customer support representatives, job recruiters, travel agents and the like.
At a showcase last year, Salesforce introduced Agentforce, a platform that allows businesses to create their own AI agents. Soon after, its rival Microsoft announced its own plan to let businesses do the same with Copilot Studio, the tech giant’s comprehensive platform for building AI agents.
Many in tech circles have already documented the technology’s growing capabilities. Aaron Levie, CEO of the cloud computing company Box, posted a video to X last month showing how OpenAI’s Operator could take the initial steps necessary to exchange sensitive documents with another company.
AI agents are a step beyond the capabilities of AI chatbots, which fueled the generative AI boom that began with the launch of OpenAI’s large language model ChatGPT in November 2022. LLMs like ChatGPT are predictive text models, meaning they generate responses to human input but don’t take independent actions.
In recent years, OpenAI and tech titans like Google and Meta, along with leading AI startups like Anthropic and Stability AI, have raced to develop multimodal LLMs capable of generating output in visual mediums like images and video.
But developers today are also increasingly integrating LLMs into a growing frontier of AI-powered tools: agents that can make decisions and take autonomous actions on behalf of their users.
The burgeoning accessibility of this technology means people can use various agents to automatically transcribe and summarize work meetings, book a flight, sign up for an exercise class, serve as a health coach or even manage their stock portfolio.
“A big question, as we as individuals use more and more of these tools, is: Do we lose critical thinking capabilities over time?” Madsen said. “We can automate some routine tasks, but then does that extra capacity in our brains now get used in another way so that we continue to advance our knowledge, our critical thinking? Or do we become relaxed and reliant on the tools?”
AI assistants and agents even crept into this year’s star-studded Super Bowl ads, such as when Salesforce had Matthew McConaughey tout its AI agent’s ability to book restaurant reservations.
“The big impact on society has both positive and negative aspects,” said Param Singh, a professor of business technologies and marketing at Carnegie Mellon University. “If you automate so much of these tasks, you are going to be more efficient as a company. But these tend to be the tasks for entry-level jobs, and those kinds of jobs are going to get taken over by AI a lot more.”
But he said that at this stage, human workers will still be needed to supervise and guide automated agents, which can’t be relied upon to perfectly interpret instructions or to understand necessary context on their own.
These tools can also be prone to mishaps — like when a driverless Waymo robotaxi drove its customer in circles instead of taking him to the airport — that are unlikely to happen under the control of human workers.
Singh said that AI agents work most dependably for repetitive tasks done in a controlled environment, where they are only permitted to make a small set of decisions and any errors would not lead to critical consequences.
“If it’s a high-stakes decision, then we need to be very careful, because even one mistake can have a big impact,” he said. “If these [AI agents] are in any kind of even slightly dynamic environment, then I don’t think we can trust them at all at this time.”
Despite their ballooning popularity in recent years, AI agents have been evolving for some time. Voice assistants like Siri and Amazon’s Alexa, as well as smart home devices like Google Home, were some of the earliest AI assistants to have been adopted by the masses. But even these devices, which could process voice commands and execute tasks, lacked the full autonomy and decision-making capabilities of today’s AI agents.
As these tools grow more sophisticated and further infiltrate daily routines, some experts say people risk becoming overly dependent on AI-powered help to usher them through life. Others are more optimistic, saying that the adoption of AI agents for simple tasks could make critical thinking and creativity all the more important.
“Recognizing that it’s a collaboration between you and the agent is a good starting point,” Madsen, the professor from Santa Clara University, said. “You are the one that should be guiding the agent to some extent, as opposed to the agent guiding you.”
