A pattern from coding agents
Coding has changed shape twice in a short span. In the IDE era you typed everything yourself; the machine helped with syntax, but the thinking and the execution were both yours. Then Copilot arrived: you described what you wanted and it filled in the code, though you still read every line before it counted. Now we are in the agent era. You ask it to refactor an API from REST to GraphQL, and it reads the codebase, plans the steps, edits the files, runs the tests, and hands back a result. You supply the intent; it handles the middle.
The strange part is what happened to the interface. It got simpler, not richer. We went from elaborate IDEs back to a plain terminal — not because a terminal looks better, but because once the system does the work, you no longer need to watch it being done.
The same shift is coming to health
Health tracking has been moving along the same line.
| Era | Interaction | Human role |
|---|---|---|
| Manual | Data points | Do everything |
| Wearables + apps | Dashboards + scores | Interpret and decide |
| Health agent | Goals | Set direction, verify |
Data collection is a solved problem. Rings, watches and bands quietly capture heart rate, HRV, temperature and sleep stages without anyone thinking about it. What we haven’t solved is the decision. Every morning still runs the same loop: open the app, check the score, read the charts, decide what to do about it. The machine measures and you think — which is exactly where coding sat before agents showed up.
What a health agent actually does
The real change is in the unit you interact with. Coding moved from characters to code blocks to intent; health is moving from data to scores to goals. A health agent in that sense isn’t a dashboard you read. It’s a closed loop: it senses, analyzes, decides, and acts.
In the morning, instead of checking your sleep score, interpreting your HRV and deciding whether to take it easy, you get a single line:
Deep sleep improved, but your HRV dropped after 3 AM, probably the late dinner. Try eating before 6:30 today — I’ve set a reminder.
No app, no chart, no decision left for you to make.
Through the day it orchestrates rather than reports:
- 10:00 AM — sustained stress, so it starts a two-minute breathing exercise
- 3:00 PM — low activity and an open calendar slot, so it suggests a walk
- 9:30 PM — bedtime approaching, so it dims the lights and switches on do-not-disturb
The biggest difference is that you can hand it a goal instead of a question. You rarely need to ask what your sleep score is. You say you want ninety minutes of deep sleep a night within a month, and it studies your history, designs interventions, runs them as small experiments, and adjusts based on what actually worked.
The interface probably isn’t an app
In coding, the agent era traded the IDE for a command line. In health the equivalent is nearly as drastic: almost everything becomes ambient — a notification, a haptic tap, the lights dimming on their own. A little of it is conversation, the occasional “why was my HRV low last night?” Only rarely do you open a screen at all.
The app turns into a debug console. You reach for it when something looks wrong, or when you’re simply curious. The next generation of these products won’t win on prettier dashboards or cleverer scores. They’ll win on whether they can get rid of the dashboard entirely.