Agentic Entropy · Eric Turkington

Like many people in technology in 2026, I build AI agents for a living, while using them to help me earn my living.

Each day I work to shape the development of automated workflows in multiple clinical categories — acute care, primary care, and national-scale interoperability. It’s a beautifully rich canvas to solve problems, get creative and push the boundaries of what these systems can do. If wielded well, agents are clearly the most capable instruments to rapidly create value in a multitude of forms, in healthcare and nearly every industry imaginable.

Value, but with massive externalities we must manage. Because I am in the proverbial engine room of this transition, I can see the exhaust we are generating. We have to manage this evolution thoughtfully, or we risk triggering a profound, systemic decay. I call this risk Agentic Entropy.

In thermodynamics, entropy is the inevitable slide toward disorder within a closed system. But in machine learning and information theory, entropy is a measure of surprise — how unpredictable a piece of information is.

The Agentic Entropy I see looming plays on a dangerous paradox between the two. On the surface, it isn’t the cinematic crash of a rogue superintelligence. It is the silent, comfortable dissolution of our own control. It is the steady erosion of shared understandings forged through imperfect, inefficient dialogues, brought on by the very tools we are building to connect and optimize us.

Bypassing each other entirely

We are rapidly approaching a reality where unmediated human-to-human interaction becomes scarcer, at least outside our innermost circles. For speed, for “quality,” we are inserting silicon proxies between ourselves and the rest of the world. We already likely receive AI inputs, and our first instinct is to run them through our own lightly prompted AI analysts, shallowly digesting that assessment and flinging it back, often for another agent’s own cognitive power to interpret and run with.

Imagine the trajectory of a simple workplace exchange in the near future. I draft a rough, unfiltered thought about a strategic pivot. My personal agent smooths my raw text into a diplomatic, homogenized memo. It lands in your ecosystem, where your agent instantly distills it back into three easily digestible bullet points for you to read between meetings. We both feel incredibly productive. We feel efficient.

But what actually transpired? We successfully bypassed each other entirely.

The nuances — the slight hesitation in my original phrasing, the specific vocabulary that betrayed my underlying anxiety about a little detail — was sanitized away by my agent, and whatever was left was stripped out by yours. We exchanged data, but we generated zero mutual understanding. Whatever potentially valuable signal there was, detectable to a human’s emotionally-tuned, nuance-appreciating brain, was buried to machine eyes in the apparent noise. And forever lost.

What if we pull that thread further? What if our businesses are run by leadership teams who only ever consume the agent-synthesized summaries of agent-generated reports? Better yet, what about businesses run by agentic CEOs? Intellectual growth and progress demands the messy, inefficient collision of different perspectives. When working through a complex conceptual problem, the breakthrough rarely comes from a frictionless transfer of data. It comes from the pushback of a colleague, the debate over a contradictory data point, the shared struggle to align the minute details with the big picture. Agent-to-agent automation strips this away. It smooths out the edges of our interactions until there is nothing left to grip.

Paradoxically, while we are sanitizing the surprise out of our human interactions, we are incubating it in the machine layer. In the near future, agents will be frequently negotiating, misinterpreting, and re-interpreting with other agents at speeds we cannot monitor and with rationales that are not easily interpretable. They will inevitably generate highly complex, emergent behaviors. If we take ourselves out of the loop in the name of efficiency, we will be entirely unprepared for the sheer scale of the surprises these autonomous networks can produce.

The clinical case for friction

To see the high stakes of this intermediation, we need only look to the clinical frontier. In my work, the market temptation to fully automate healthcare communications is immense. But consider the doctor-to-doctor handoff. When an anesthesiologist steps away from the OT to brief the PACU team, they are transmitting far more than vital signs and procedural codes. They are transmitting a narrative. They are passing along body language, gut feelings generated by accumulated experience, and clinical instincts not fully available to conscious explanation. These are the vital nuances that defy structured data capture.

This is precisely where the agent shifts from a proxy to a partner. Voice-powered AI is infinitely better suited to track the raw, minute facts of the case, instantly contextualize them against a vast patient history, and ensure that structured data is front-and-center at the moment of transfer. By offloading the rigid arithmetic of the patient’s record to the machine, we free the anesthesiologist to remain entirely present. We preserve the receiving doctor’s ability to ask a clarifying question, or catch a subtle change in tone indicating urgency. We keep the shared reality that catches edge cases before they become emergencies.

The agent manages the volume of the data; the physician manages the value of the care.

Designing for human collision

Zooming back out to society at large, this dynamic is playing out on many scales. We are outsourcing discernment to networks of agents that optimize for speed, politeness, and narrow parameters, fundamentally altering how we relate to one another. We are becoming isolated nodes, insulated from the challenging, vitalizing reality of other people’s unfiltered minds. We need to experience the good, the bad, and the ugly of one another. The entropy here is the loss of our collective capacity to truly understand differing perspectives, because we are no longer actually interacting with them. Or because they are never truly being formed in carbon-based neural networks.

The antidote to Agentic Entropy is not to reject the technology. I welcome AI to clear the administrative static that burns us out and clogs our calendars. But we must radically rethink its boundaries. We have to apply a higher level of practical rigor to our product designs and our own lives, recognizing that convenience is not a synonym for connection, and that efficiency should never come at the cost of shared understanding.

If we are to navigate this transition without losing the essence of what makes our collaborations meaningful, we have to intentionally design for human collision. We have to train AI systems that preserve the uniqueness of our personalities, the imperfections of our thoughts, the half-baked threads that get pulled by others into brilliance. We must aggressively protect the spaces where we encounter each other directly, embracing the un-sanitized, human frictions that ultimately keep our systems — and our society — in balance.

This piece first appeared on LinkedIn, February 2026.