# DiabAgent: Toward Agentic Physiological Control

Source: [arXiv](https://arxiv.org/abs/2611.03601v1)  
Feed7 permalink: https://feed7.dev/p/2611-03601v1-03z1nfj  
Published: 2026-11-02T18:41:02.000Z  
Trust: Needs Review (needs_review)

## Why Included

DiabAgent evaluates sustained agentic glucose control with a simulator and 212-task benchmark. In the reported setup, its agent framework scored above the OpenAPS baseline on COGI.

## Source Summary

DiabAgent combines the **APC framework**, DiabetSim simulator, and SweetBench benchmark. SweetBench instantiates **32 scenario designs as 212 tasks**; with Claude Opus 5.5 and hourly decisions, APC scored 73.5 COGI versus 71.2 for the OpenAPS baseline.

## Practical Implication

For builders of long-running tool agents, the setup highlights per-user adaptation, accumulated experience, and context such as meals or exercise. The authors also explore combining frequent rule-based control with less frequent agent decisions.

## Agent-Ready Context

DiabAgent combines the **APC framework**, DiabetSim simulator, and SweetBench benchmark. SweetBench instantiates **32 scenario designs as 212 tasks**; with Claude Opus 5.5 and hourly decisions, APC scored 73.5 COGI versus 71.2 for the OpenAPS baseline.

For builders of long-running tool agents, the setup highlights per-user adaptation, accumulated experience, and context such as meals or exercise. The authors also explore combining frequent rule-based control with less frequent agent decisions.

This is a virtual research evaluation for type 1 diabetes control. The supplied material does not establish safety or effectiveness in real clinical use.

## Connected Context

Feed7 judgment across 990 accumulated Signals:

DiabAgent adds a simulated, personalized control benchmark with explicit experience and context, and compares its agent to an established controller. The result supports evaluating agents on sustained decisions in a domain-specific simulator, while the virtual setting leaves real-world clinical safety unresolved.

- [Simulated Control Evals](https://feed7.dev/p/2610-31593v1-1lanp1p) — Both evaluate consequential agent control in simulated environments; Simulated Control Evals specifically cautions that simulated results are deployment-specific and do not replace executed safety evidence.
- [Patient-specific rubric-based evaluation of large language models across longitudinal cancer care](https://feed7.dev/p/2610-35847v1-0se27hb) — Both concern clinical agent evaluation across context-dependent decisions; the cancer-care benchmark shows why aggregate performance may obscure failures that vary by workflow phase.
- [AgentHPOBench: A Benchmark For Evaluating LLM Agents as Sequential Hyperparameter Optimizers](https://feed7.dev/p/2607-29626v1-1dfg6xz) — DiabAgent emphasizes accumulated experience and per-user adaptation, while AgentHPOBench tests whether agents improve through sequential history; together they point to evaluating sustained behavior, not only isolated outputs.

## Context Map

- Layer: benchmark
- Domains: research, data
- Topics: agent-evals, agent-reliability, tool-use

## Uncertainty

- This is a virtual research evaluation for type 1 diabetes control. The supplied material does not establish safety or effectiveness in real clinical use.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.