The application of agentic AI in life sciences reached a tipping point in 2025. Frameworks where specialized agents collaborate on tasks, sharing work, dividing problems, communicating with each other, went from an emerging topic to the predominant conversation. For many in the industry, this felt like a breakthrough.
The adoption was new but the thinking behind it isn’t. The foundational ideas behind agentic architectures were laid out forty years ago by Marvin Minsky in The Society of Mind. Minsky described intelligence as collections of small agents and agencies, working together to solve problems that none could handle alone. The interaction between the parts leads to solutions. In other words, agentic AI. The world rediscovered it last year.
What took so long? Partly it’s just the nature of how ideas move from research to practice. There’s an old saying in the field: the last 10 percent is 90 percent of the work. You think there’s only 10 percent left, so the solution feels right around the corner. But that remaining stretch is where most of the time and effort actually goes. It creates a gulf in expectations about when things are going to happen. I do research at Sorcero, so I’ve lived with that gulf for a while. Forty years, in this case.
But the timeline alone doesn’t explain it. Something else happened along the way.
What the industry rediscovered
For a lot of people, the conversation about AI in life sciences had narrowed into a conversation about large language models. LLMs became synonymous with AI itself. Anyone who actually works in the field knows that’s not true. Alan Kay made this point during his keynote at the “75th anniversary of the Turing Test” celebration: some of the most important AI-driven advances in life sciences, such as AlphaFold, Google’s protein folding project, have nothing to do with LLMs.
The recent fixation on LLMs has meant that other useful AI technologies have been back of mind. This attention is merited: LLMs (and their application to Generative AI) are remarkably powerful. But in embracing this “shiny new thing,” we lost sight of other useful work. With Agentic frameworks, we’re able to broaden our vision again. And the organizations that can see past the hype—and past a single model architecture—are making advances in the marketplace faster than everybody else.
The trust problem
The biggest gap in the AI industry’s thinking is still this: Gen-AI, applied on its own, is incomplete. That matters in life sciences more than anywhere. In life sciences, you have to be able to trust the output. There is no getting around it.
Life sciences is evidence-based. The only way you trust anything in this field is if it’s grounded in evidence, and that evidence is typically in the form of a clinical trial. Gen-AI is good at what it does, but what it does is incomplete. LLMs don’t do grounding. You can get a result. But you don’t know that you can trust that result.
Think of Gen-AI like surgery. Think of the surgical team as a collection of agents working towards a common outcome. And what happens in the operating room (OR), by analogy that’s the Gen-AI part. But the success of the surgery—the outcome that matters—also depends on what happens before the patient enters the OR (pre-op) and after they leave the OR (post-op). The tests and planning before and the monitoring and care provided after are critical.
In the context of Gen-AI, the data modeling work that happens before and the verification work that happens after are necessary to ground the application of the model and to build trust in the output. Some industries can get away with ambiguity in inputs and outputs. Life sciences cannot.
Only once you’ve done the grounding work and verified that what you’ve built is trustworthy can you start to consider efficiencies.
From efficiency to reach
We tend to sell AI in life sciences in terms of its efficiency: doing things faster and cheaper. A savings in both time and money. But there’s another way of thinking about efficiency. You can use it to expand the reach of your message: more targeted storytelling for more audiences.
Subgroup analysis is a good example. In clinical research, subgroup analysis means looking at how a treatment affects specific populations, defined by age, gender, or underlying conditions, rather than reporting just the overall result. Reporting subgroup analyses are becoming increasingly important in life sciences, but doing it is expensive. And because it’s expensive, we don’t do enough of it.
One example is a plain-language summary, which regulators require to tell the story of a clinical trial to study participants. A plain-language summary is typically a single document catering to all participants, regardless of their subgroup.
But the study population—the audience for the summary—can be quite diverse. Different demographics, different reading levels, and different languages. When journals were distributed in print, I understand why you’d only want one summary. Print real-estate is expensive. But today’s journals are universally online. Page counts don’t impact costs. So why have just one summary?
This is where agentic AI in life sciences starts to unlock real value beyond cost savings. GenAI makes it feasible to tell more stories, to more audiences, in more formats. A plain-language summary at every reading level and in every language. Or video for people who’d rather watch than read. And tailored to specific subgroups, because the adverse events that affect one population and the benefits that matter to another aren’t the same.
Last year I did a small pilot around telling the story of vaccine trials to vaccine skeptics. That’s a subgroup that really needs to be reached, and yet we struggle reaching them with the generic story. When we tailor the story, the evidence doesn’t change; it’s still grounded in the clinical trial. But we do a better job of telling that story to that particular audience.
Years ago—when Minsky was inventing his society of agents—I quipped that Marshall McLuhan was wrong when he said “the medium is the message.” I argued at the time that “the message is the message” and we can exploit different media to convey the message.
Today, in the life sciences, we have a message derived from the evidence. We have the opportunity to relay that message through many different media. To tell it to this audience through this medium, that audience through that medium. While the media might dictate to some degree how we tell these stories, each story comes from the evidence.
This is where I think we’re heading: we can reach a lot more people, a lot more effectively, through Gen-AI and agentic frameworks. And to me, that’s worth paying attention to. It only took us forty years to get here.
About the Author
Walter Bender is Chief Scientific Officer at Sorcero and former executive director and senior research scientist at MIT’s Media Laboratory.
Disclosure: The author is employed by Sorcero, which develops AI-powered tools for life sciences communications.
