Posted on https://moseshng.wordpress.com/2026/06/17/the-moving-car-dilemma/ What if every citizen had an AI agent for their entire life? ## The Moving Car Dilemma I'm currently helping implement a digital mental health platform — the kind that uses passive sensing to pick up early signals so a case worker can proactively reach out before things spiral. On paper, it's a clean idea. Catch the dip in sleep or mobility or speech before the crisis call comes in. In practice, something quieter and more stubborn shows up: that monitoring is additional work. Somebody has to look at the dashboard. Somebody has to decide what the dip means, and then make the call. The system doesn't remove labour. It just relocates it to people who already have none to spare. I keep turning this over, partly because the same logic doesn't stay contained to mental health. Primary care already manages a caseload in the millions. The idea of layering passive sensing on top of that — asking already-stretched clinics to also become interpreters of continuous behavioural data — doesn't survive contact with how those clinics actually run. There's a framework health economists use called the Iron Triangle: access, quality, cost. Pick two. It's a neat model. Almost too neat. Once you see it, you start noticing how every policy decision quietly bends toward two corners while straining the third. Expand access, and costs rise. Cut costs, and something — time, attention, quality — gives way. Try to preserve quality at scale, and the system slows under its own weight. Singapore has been unusually disciplined about balancing this triangle. Co-payments to temper overuse. Tight governance. Aggressively organised primary care. For a long time, it worked. But the maths is shifting under it. The proportion of citizens aged 65 and above has gone from 13.1% in 2015 to 20.7% this year, and the national healthcare budget has just crossed $21 billion. The pressure on the triangle isn't increasing — it's compounding. More chronic illness, more follow-ups, more coordination, more time. The system isn't breaking. It's stretching in ways that are getting harder to absorb. And this is where most proposed solutions start to sound the same: digitise, monitor remotely, shift care into the community. Group doctors and care coordinators into "teamlets" to manage chronic disease at scale. On paper, it all makes sense. In practice, it runs into the same wall I keep running into. Every new layer of "smart" monitoring quietly assumes someone has spare time to look at it. A dashboard gets built. Alerts get generated. Data streams in. And the clinician — already moving from patient to patient, task to task — is now expected to also be an interpreter of continuous data, a proactive coordinator, a remote sentinel. The obvious answer is to just hire for it — carve out a dedicated role whose entire job is watching the dashboard and making the calls. But that's not really an escape from the triangle, it's the triangle showing up again under a different name. Headcount is cost, and cost is the corner that's already buckling. A new layer of staff doesn't scale at the rate the data does, and the same workforce crunch that makes it hard to fold monitoring into existing roles makes it just as hard to hire a fresh team to do it instead. It feels like trying to rewire a car while it's doing 140 km/h on the highway. The engine's hot, the road is full, there's no shoulder to pull into. You can't hand the driver a blueprint and ask them to rethink the machine mid-journey. So the question I keep landing on: what if the redesign doesn't happen inside the car at all? What if it happens alongside it? Imagine that instead of pushing more responsibility into clinics and hospitals, you build something parallel — a layer that follows the person, not the institution. Not one giant "AI doctor" (which hits context limits, hallucinates under pressure, and becomes a single point of failure the moment it's wrong), but a constellation of small, narrow, specialised agents, persistent over a lifetime, coordinated by something sitting above them. Some of it is almost boringly administrative — an agent that just handles logistics, chasing appointments across institutions, rebooking missed slots, pulling the right lab results from the national record before a referral goes out. Zero clinical judgement required, but it claws back a genuinely enormous amount of human time. Above that sit the clinical specialists, operating inside locked, guideline-bound sandboxes. One tracks medication adherence — smart blister packs, refill timing, contraindication checks — and quietly arranges the pharmacy delivery before supplies run out. Another acts as a low-intensity therapeutic companion: not a replacement for a clinician, but a daily, always-available presence running brief behavioural check-ins and tracking how someone responds over time. And underneath all of it, the part closest to what I actually work on: something pulling quiet streams off a wearable and a few sensors around the home — sleep, heart rate variability, how often someone actually leaves the house — and watching for drift, not in a dramatic or diagnostic way. Language thinning out, typing changing speed, the smaller things too. It doesn't diagnose. It maps deviation from someone's own decade-long baseline. When enough of those deviations stack up, a triage layer pulls everything together — the missed doses, the falling HRV, the spike in distress scores — into one compressed brief, and hands a human exactly two buttons: Patient, 68. Phenotype agent: 40% drop in out-of-home mobility over 14 days, HRV trending down. Pharma agent: 4-day medication lapse. Therapy agent: rising distress on check-ins. Recommended action: trigger pharmacy delivery, flag care coordinator for a call. Approve / Reject. The clinician never wades through the raw noise. They don't monitor, they don't chase — they get handed a decision already built, and they make the final call. That's the whole shift: less surveillance work pushed onto people who are already at capacity, more decisions handed to them pre-packaged. The car keeps moving. Something else runs beside it, watching the road the driver doesn't have time to watch. ## Why cannot? And then, almost as soon as I've built the idea up, it starts to strain. How accurate does something like this need to be before anyone should trust it with a life — and what does a 1% error rate actually mean when the missed signal is a suicide risk rather than a missed appointment? What does it mean to be watched this continuously, even when the intent is entirely caring? There's a real gap between someone tracking their own steps and someone consenting to a permanent, passive record of their voice, movement, and typing. Cross that line without enough trust built first, and people don't comply — they either opt out or start gaming the system, which quietly poisons the very data the thing depends on. Can an architecture, a data model, a set of baseline weights established this year actually survive technologically intact across someone's entire life, decades of platform shifts later? Software doesn't usually get to assume that kind of continuity. What happens the day this gets breached — a single store holding someone's entire psychological and medical history, end to end? That's not a normal data leak. That's the kind of failure public trust in healthcare infrastructure doesn't recover from quickly. And underneath all of it, the blunt question: can a population of six million actually afford to run agentic loops like this, continuously, for everyone, without the compute cost quietly eating the very savings the whole idea was supposed to produce? These aren't edge cases sitting at the margins of the idea. They're the core of it. Which is maybe why I don't think of this as a proposal. It's more of a direction — a way of asking where the burden of care might sit, once the current arrangement stops being able to stretch any further. Right now, we keep asking the system to absorb more. The clinic stretches. The case worker stretches. The dashboard gets built, and someone who was already full finds a way to also watch it. Maybe the more interesting move isn't finding a way to stretch it less. Maybe it's building something next to it instead. ## Maybe change the road? not the car? Or maybe I'm still arguing about the wrong layer. A faster car, a parallel car, even a car that needs no monitoring at all — none of it answers why so many people end up needing the car to begin with. There's a newer instinct in Singapore's healthcare policy aimed at exactly that question — anchoring residents to a family doctor early, betting that catching the drift before it becomes a diagnosis matters more than treating it well after. That's not a faster car. That's a different road, built so fewer people ever need the emergency lane at all. It's tempting to think the upstream version might solves what the downstream version couldn't with a personal citizen AI coach tracking diet, sleep, lifestyle, movement, mood from young. But it's the same AI agent, just doing an earlier job. The hallucination risk doesn't shrink because the stakes feel gentler. The surveillance question gets harder, not easier, when it starts at sixteen instead of sixty-eight and runs twice as long. The data still has to survive, intact, across an entire lifetime of platform shifts. The breach is still catastrophic — just earlier, and longer-lived. I don't know la. Wait for smart people to solve it la. Just brain farting la. -- Related - [[Open health - Getting Singapore’s healthcare fit for the future]] - [[20241002 AI Horizons - Exploring the Future of Innovations]] - [[20241002 Medical Informatics and AI shaping our future - Clinical and Industry Perspective]] - [[20241002 Deploying Generative AI in Clinical Environments]]