Agentic artificial intelligence took center stage at the HLTH 2025 conference in Las Vegas, and Brigham Hyde, Ph.D., is bullish about the future of the technology in healthcare.
Atropos Health, which Hyde co-founded, is focused on filling the "evidence gap" in healthcare by pulling relevant clinical data and generating real-world evidence at the point of care. At the HLTH conference, the company unveiled an agentic AI solution, called Evidence Agent, a personalized real-world evidence generator that integrates directly into electronic health records and clinicians' workflows.
Stanford Health Care has deployed Evidence Agent, which works with the health system's ChatEHR, an AI tool that translates patient records into plain language, and Microsoft's Dragon Copilot.
Atropos' Evidence Agent accesses the patient record and patient level information such as “reason for visit” via ChatEHR and ambient data from Microsoft Dragon Copilot. Using the data, Evidence Agent generates patient-specific RWE in order to provide documentation for treatment options or other care decisions without the physician ever needing to ask a question, according to the company.
The RWE from the agent is integrated into the EHR, allowing the physician to save time by staying in the workflow.
“Utilizing agentic AI to embed RWE into workflows with patient-specific insights will enable providers to proactively address patient questions with speed and accuracy,” said Michael Pfeffer, M.D., chief information and digital officer at Stanford Medicine, in a statement. "This example shows how ambient AI, ChatEHR and agentic AI are putting technology to work for the benefit of both providers and patients.”
While many health tech companies can summarize clinical research at the point of care, Atropos Health answers questions with both existing evidence and offers the option to generate new evidence. In 2023, Atropos launched a new operating system, Geneva OS, and a chatbot interface to help generate observational studies rapidly and at scale. That technology, ChatRWD, reduces the time to produce high-quality publication-grade RWE from months to minutes through a chat-based AI co-pilot, according to the company.
Through Atropos, doctors and researchers can access 100,000 novel studies to find more answers to clinical questions not available in existing published literature, according to Hyde. The company also announced last week that it was making its platform available to U.S.-based clinicians for free as part of a no cost trial.
"Our technology installs behind the health system's firewall, so no data leaves and it's all de-identified. Once we have the data, you ask your question, and your question might be something like, 'For diabetic patients, what works better, GLP-1 or Metformin? But I'm only interested in liver disease patients. If you ask a solution like OpenEvidence that question, they'll reference a study comparing GLP-1 and Metformin, but it will highlight this and say, 'There's no study on liver disease patients' so you don't have a direct answer," Hyde, CEO at Atropos Health, told Fierce Healthcare during an interview at the HLTH 2025 conference last week.
"Whereas us, we run that study on that patient data in about 20 seconds and actually produce a study, which usually ends up being the largest patient study ever run on these questions, that directly, with a p-value and all the statistics, answers the question. We fill the gap that exists in literature. We call it the evidence gap. There's just not enough studies. We built a machine to automate the production of high-quality studies," Hyde said.
The company was founded in 2020 as a spinout of the Green Button technology at Stanford University and developed a consultation service for doctors powered by publication-grade RWE to guide clinical decisions.
Stanford Health Care has been working with the company to integrate RWE into physicians' workflow and clinical notes including by integrating with ambient AI tech and its EHR.
Atropos' RWE agent is now available to all U.S.–based health systems for integration into their physician workflow, the company said.
"We built the Atropos Evidence Agent with the same rigor we apply to all our solutions prior to deploying within a health system,” Hyde said. "We have strong guardrails in place to ensure quality and accuracy for the patient specific personalized evidence generated by the Atropos Evidence Agent."
Atropos has broader ambitions to integrate its Evidence Agent into ambient AI workflows. As part of a pilot with Microsoft, Atropos will combine ambient encounter data from Dragon Copilot with its agentic AI solution.
“We are in a moment in the technology industry where AI is helping to address many of healthcare’s most urgent problems," said Peter Durlach, CVP, chief strategy officer, Microsoft Health and Life Sciences, in a statement. “By collaborating with Atropos Health, we are working to help healthcare providers unlock new levels of productivity, reduce administrative burden and empower them to leverage new clinical decision making capabilities in the care of their patients.”
Agentic AI represents a potential major shift in how physicians and clinicians interact with healthcare technology solutions, or what Hyde refers to as a "shift in glass."
After the iPhone came out, many people stopping using their laptops and shifted to using their iPhones, he noted.
"What agentic AI represents is 'Why do I need to open an app if I can just talk to an agent in my workflow?' Or even in the ambient case, it's just doing the work for me. I think this is bigger than just us. Why do we need to log into Epic anymore if I'm recording the visit and I can talk to an agent?" Hyde said. "I think Microsoft is in the leadership position there."
In May, Microsoft rolled out an agentic AI orchestrator, a new product that coordinates AI agents for multidisciplinary care teams. Its initial framework is designed to study the healthcare AI agent orchestrator for assisting tumor boards, a group of multispecialty oncologists that review tumors to determine treatments for patients. Cancer teams at Stanford University are using the agent orchestrator as are teams at Johns Hopkins, Providence Genomics, Mass General Brigham and the University of Wisconsin School of Medicine and Public Health.
"In a world where physicians can just invite agents into a chat, talk to them and get their work done, that could be a shift in glass," Hyde noted.
He added, "We're super excited about where agents can go."
Many health tech companies are looking to build AI-driven solutions that give providers faster access to data for clinical decision support. Atropos' ability to deliver full observational studies sets it apart, Hyde asserts.
"We've always been about converting healthcare data into high-quality evidence at scale," Hyde said. "Interestingly, in this part of the cycle where the [large language models (LLMs)] are all competing, it turns out having novel evidence is a training advantage for an LLM. You can answer questions better and more questions. If I were thinking about our big trajectory, we've got about 100,000 novel studies right now that nobody else has. The back-of-the-envelope math for us is that we can be larger than PubMed within three years. So, imagine the entire corpus of medical literature being replaced and expanded upon by techniques like ours, and what that means both for more evidence for care and for LLMs."
Atropos' novel, methodologically credible RWE reports are filling evidence gaps in specialty areas like geriatrics where there is a lack of new clinical evidence, as well as hematology and oncology, Hyde noted.
As AI tools are rapidly adopted in healthcare, clinicians need to ensure the answers they receive from AI-powered solutions are accurate and trustworthy.
"For physicians to trust something, they have to know they're going to get a high-quality answer every time," Hyde noted. Atropos developed a simple badge system—based on green, yellow and red—to provide users with clear feedback on the reliability of a LLM-generated response to a clinical question. A green badge signifies a high-confidence, evidence-grounded response, according to a paper Atropos published.
"Most LLMs provide you information that might be relevant, but don't clearly say whether they've answered your question, and I think that erodes trust. We're trying to put front and center, 'Did this answer your question with evidence that you trust and trust enough to make a treatment decision?' I think that's critical," Hyde said.