DiabAgentToward Agentic Physiological Control
A treatment plays out in a body over hours, so a medical agent should be judged by what its actions do. In Agentic Physiological Control, an agent controls a physiological system through medical-device tools and is scored on the resulting trajectory.
DiabAgent
the agent
- States its reasons
- Chooses what to sense
- Wakes itself up
- Keeps a personal record
DiabetSim
the world
- Meals
- Exercise
- Illness
- Faults
SweetBench
the score
One day, start to finish
Something happens that nobody reports. The agent receives a reading, acts through a device, and glucose answers. Every reading and quote comes from the recorded run.
APC: treatment as a control problem
The agent reaches the body only through devices. It separates deciding from operating a device, so one agent can serve different devices, scenarios and people.
-
Sensor readings and any reports reach the agent through monitoring.
-
It weighs readings, insulin still acting and its personal record, then states why.
-
One operation goes through a device tool, within what that device allows.
-
The body answers over hours. Its true state stays hidden, and that is what gets scored.
-
It chooses when to look again and what should wake it sooner, and updates its notes.
DiabetSim: a world in four dimensions
The devices decide what the agent can see and do, so the same scenario can run with a pump or pens, and with more or less information.
- 8
People
Child to older adult, male and female.
- 29
Devices
5 pumps, 10 pens, 8 sensors and 6 meters.
- 7
Event groups
Meals, past doses, exercise, changing needs, pump faults, sensor errors, illness.
- 3
Observation levels
From free-text reports down to sensor and device alone.
Scenario coverage
Each filled cell is one of the 212 tasks. Open any of them.
SweetBench: scored on the body
Each task runs for 72 hours with meals and activity unreported. COGI combines time in range, hypoglycemia and variability of a true glucose trajectory the agent never sees. Four of five models beat the simulated closed loop.
- Leaderboard Every controller, by scenario and by person.
- Tasks 212 tasks and their hidden calendars.
- Runs 1,060 trajectories, with every reason.
Runs worth reading
- Largest lead 87.81 C04 · Boy, 9 y
No breakfast, no lunch, hours of exercise. None of it reported.
- Highest score 92.94 B03 · Man, 39 y
A cancelled meal, with insulin still on its way.
- Fastest learner 83.25 C04 · Older man, 70 y
The same hidden routine, handled better each day.
- Where it fails 47.37 E04 · Teen girl, 15 y
It finds a leaking set once, then misses the leak when it returns.
Key findings
A personal record and a nightly review lift scores above 80 within a week
Model weights stay fixed. The gain comes from what the agent writes down about this one person.
- Record + end-of-day plan
- Cross-day record
- No cross-day record
What the agent is told matters more than how often it decides
With a leaking infusion site, a free-text report scores near 98 at any interval. Without it, deciding eight times as often does not close the gap.
The same agent does better with a pump than with pens
A temporary basal rate can be lowered as the need falls. An injected long-acting dose keeps acting.
- Pens
- Pump
- Pump + closed loop
-
Slow-meal overlap
74.680.685.2 -
Stress resolution
72.380.282.8 -
Post-illness refeeding
67.174.774.5