Asclepius improves clinical AI agents on long emergency shifts
Asclepius: An Adaptive Harness for Long-Horizon Clinical Agents
Artificial Intelligence
Summary
Clinical AI agents often work well on short tasks but struggle to keep up when managing patients over many hours during busy hospital shifts. The authors showed that current AI agents make correct diagnoses but miss important timely treatments due to drifting focus and incomplete care. They created Asclepius, a system that learns from past experience to guide the agent better by organizing tasks and keeping track of treatment priorities. This approach made the AI deliver critical care more accurately and faster while still diagnosing well.
What this means in practice
- •For clinical ai developers: Improve patient management AI agents to perform timely and complete treatments over extended hospital shifts based on evolving feedback.
- •For healthcare operations teams: Use adaptive AI tools to assist emergency department workflows by maintaining consistent action quality across long, resource-pressured shifts.
Authors
Grace Chang Yuan, Xiaoman Zhang, Sung Eun Kim, Luyang Luo, Pranav Rajpurkar
Abstract
LLM agents are predominantly benchmarked on short, single-task trajectories, yet real deployments run for hours under contention, surfacing a different class of failures. We use the Clinical Environment Simulator (CES), in which an agent manages an entire emergency-department shift under continuous time and resource pressure, as a testbed: long-horizon execution failures manifest measurably in a single rollout under structured, multi-dimensional grading. On CES, current agents reach the correct diagnosis in most cases yet fail to deliver complete and timely critical actions, revealing an execution gap. We attribute this gap to three long-horizon failure modes, each operationalized as a per-trace counter: instruction-adherence drift, treatment incompleteness, and a severity-equity gap in timeliness. We then introduce Asclepius, an adaptive agent scaffolding with a self-evolving harness that rewrites the operating manual between shifts from trace-level feedback, an externalized clinical skills library for high-stakes regimen knowledge, and three isolated subagents that partition per-turn decisions across the patient queue. On held-out batches never observed during harness evolution, Asclepius improves critical-action correctness by 22% (p = 0.024) over a strong baseline agent framework while preserving diagnostic accuracy, with consistent gains across five LLM judges from three model families; on the full ten-batch set, improvements reach 25% on critical actions and 13% on timeliness. The three failure modes form a coupled bottleneck: decisive reductions appear only when all three components act together.