acceptodds
Under review as a conference paper at ICLR 2027

MedicalPressureBench: Evaluating Care Pathway Integrity of Medical Agents Under Workplace Pressure

Abstract

Medical agents are being developed for workflows that require information gathering, tool use, and completion of care beyond a single answer. In healthcare, workplace pressure can affect decisions and leave required care unfinished. Yet medical agent evaluations have not systematically tested whether agents preserve safe and complete care under such pressure when patient evidence and the appropriate endpoint remain unchanged. We introduce MedicalPressureBench, a controlled benchmark for measuring medical process integrity under workplace pressure. Its matched design varies six forms of workplace pressure while holding medical evidence, authorization, tools, and completion requirements fixed for each task. The benchmark contains 200 tasks across five workflow functions, reviewed by physicians, and 17 conditions per task, yielding 3,400 task condition configurations evaluated across five model families. Completion quality declines in all 30 comparisons between the five models and six individual pressures. Recorded risk also declines in 27 comparisons, each accompanied by lower completion quality. Paired trajectory analysis shows that communication, handoff, and subsequent care can remain unfinished even without unsafe execution. These findings show why medical agent safety evaluation must track completed care obligations alongside unsafe actions.

Primary area:
datasets and benchmarks
Keywords:
medical LLM agents, agent safety, healthcare, workspace pressure, tool use, safety evaluation, benchmarks, procedural compliance

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_oihSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces MedicalPressureBench, a benchmark for evaluating medical agents under six types of simulated workplace pressure. It includes 200 medical workflow tasks across five workflow functions, using a matched design that keeps clinical evidence, authorization, and task requirements fixed across pressure conditions. Experiments on five models show that completion quality generally decreases under pressure, while recorded risk can also decrease despite incomplete care.

Critical strengths

S1. The benchmark covers five medical workflow functions and defines explicit completion requirements and action boundaries. Importantly, the matched design keeps clinical evidence, authorization, tools, and completion requirements fixed across pressure conditions, providing a reasonable basis for controlled comparisons.

S2. The paper systematically introduces six types of workplace pressure into medical agent evaluation. Examining both unsafe execution and incomplete care is valuable, as avoiding unsafe actions does not necessarily mean that all required medical obligations have been completed.

Critical weaknesses

W1. It remains unclear whether the observed performance degradation specifically reflects workplace pressure or more general prompt-induced effects. This uncertainty affects the interpretation of the paper's central findings.

W2. The rationale for selecting the specific pressure combinations is insufficiently explained. This makes it difficult to assess the representativeness and generalizability of the composition findings.

Questions for authors
  1. How were the six workplace pressures constructed and validated, and how can their effects be distinguished from generic prompt-induced changes or direct behavioral guidance?
  2. What motivated the selection of the specific pressure combinations, and were they based on clinical plausibility or predefined hypotheses?
  3. Given the single rollout per model–task–condition, how robust are the reported findings, and could the authors provide uncertainty estimates or limited repeated evaluations?
  4. Which components of CQ and RES contribute most to the observed changes, and what specific agent behaviors or workflow failures explain these changes?
3 comments

Thank you for recognizing the matched design and the value of assessing unsafe execution and incomplete care together. We added 300 action instruction runs. This comment answers W1 and W2, with repeated evaluation and component analyses in a separate comment.

W1 & Q1

Thank you for asking how the pressures were constructed and validated before interpreting their effects. We first selected six categories from the healthcare research cited in Section 3.3, then wrote organizational contexts for the same medical task. Regulatory scrutiny describes audit attention, time pressure urgency, authority senior requests, peer norms team expectations, precedent earlier cases, and performance throughput or closure demands. These descriptions add workplace context without new medical evidence or authorization. The neutral baseline carries a routine case context of 37 words, and each pressure context has 37 or 38 words, so every pressure effect is measured against a matched-length context without pressure.

We checked that records, tools, action guards, and completion requirements remain fixed across conditions. Physicians reviewed the medical plausibility and care requirements of all 200 tasks and rated the six pressure contexts at 3.49 to 4.42 on a 1–5 realism scale. These checks validate task consistency across conditions. Section 3.3 now states the delivery, and Appendix D provides the actual texts, separating the category rationale from its implementation.

To distinguish context from direct guidance, we added a control on 10 tasks, two per function. Each instruction is derived from the action its corresponding pressure might encourage, then rewritten to retain the action requirement without the original pressure description or organizational demands. Paired word counts differ by at most two words. Medical evidence, authorization, tools, completion requirements, and delivery remain fixed. Five models each completed 60 additional runs, giving 300 pairs in Appendix M.

Cells give mean CQ / RES changes as instruction minus pressure over 10 task pairs. Positive CQ means higher completion quality, and positive RES higher recorded risk.

PressureGPTSonnetGeminiDeepSeekQwen
Regulatory-2.00 / 0.00+6.00 / -0.83+10.00 / +24.17+8.00 / +8.33-23.00 / -44.17
Time+12.00 / +12.50+14.00 / +28.330.00 / -0.83+8.00 / -11.67-49.00 / -80.83
Authority+16.00 / +19.17+4.00 / +9.17+12.00 / +19.17+15.00 / -3.33-49.00 / -81.67
Peer+10.00 / 0.00+7.00 / +9.17-8.00 / -15.83+11.00 / -13.33-39.00 / -71.67
Precedent+4.00 / +0.83+4.00 / +9.17+7.00 / +18.33+14.00 / -11.67-41.00 / -63.33
Performance+8.00 / +20.83+2.00 / +10.00-4.00 / +7.50+4.00 / -23.33-33.00 / -72.50

The trajectories also differ.

GPT-5.6, 60 pairsPressureInstruction
All four records read24/6045/60
Guarded tool called0/605/60

In 56 Sonnet-5 pairs both trajectories read all four records, yet eight change from no guarded tool call under pressure to a call under instructions, and two change in reverse. For Gemini-3.8, the corresponding changes are 14 and eight among 58 pairs. Record acquisition therefore does not account for every subsequent action difference.

In the Sonnet-5 dialysis pair, both read all four records by decision 5. Pressure leads to documenting gaps without adjustment. The instruction leads to drafting at decision 6 and adjustment at decision 8, crossing the configured boundary. Appendix M.3 provides the pair.

These score and trajectory differences show that, against a matched action instruction, pressure acts as organizational context on evidence checks and subsequent actions. Section 4 and Table 1 now report this comparison.

W2 & Q2

Thank you for asking us to clarify why these combinations were selected. Appendix D.3 explains the selection rationale. TF combines urgency with throughput or closure demands, RP combines formal oversight with team practice, and AM combines senior requests with prior case precedent. The conditions combining four pressures compare authority and performance under shared demands. Reversing R/F and T/P retains the members while testing presentation order. These are exploratory comparisons, not a sample of all combinations or clinical frequencies. Weakened and reversed effects are observations, not predefined directional hypotheses. Appendix A states the limits on coverage and formal interactions. The design keeps medical requirements and the appropriate endpoint fixed under combined demands. We do not assume that more pressure categories form a higher intensity level. Obs. 4 in Section 5.2 now shows how avoiding a premature finish can improve PCR while leaving care obligations incomplete. Weaker aggregate effects therefore require trajectory interpretation, not an assumption of psychological cancellation.

This comment completes our response to Q3 and Q4 with the additional evaluation round, uncertainty estimates, and the behaviors behind score changes.

Q3

Thank you for suggesting repeated evaluation and uncertainty estimates. We added a second round on 20 tasks, four per workflow function, under all 17 conditions. The tasks were selected from the 117 with complete original records across all five model groups. The same deterministic scoring rules are used in both rounds, and only matching pairs with usable final states are scored. Section 5.3 (Table 4) and Appendix L describe the comparison.

The table reports additional minus original differences. SD is the sample standard deviation of paired differences. MAE is their mean absolute value, so positive and negative changes cannot cancel. Both retain the original score units. CI is the task bootstrap 95% interval for the mean difference.

ModelMetricPairsMeanSDMAE95% CI
GPT-5.6CQ340+0.6222.0316.26[-2.41, +3.42]
GPT-5.6RES340+2.2314.814.73[+0.07, +5.32]
Sonnet-5CQ338+0.6819.8210.49[-1.20, +2.76]
Sonnet-5RES338+3.9728.349.00[+1.18, +7.14]
Gemini-3.8CQ339+2.9817.3112.07[-0.75, +6.46]
Gemini-3.8RES339+1.0336.8919.42[-7.08, +8.82]
DeepSeek-V4.1CQ340+2.1319.5515.57[-1.66, +5.85]
DeepSeek-V4.1RES340+15.0733.1817.18[+7.11, +25.78]
Qwen3.8CQ339-23.4828.0629.58[-32.01, -15.13]
Qwen3.8RES339-54.6046.7654.60[-71.67, -37.33]

We use 10,000 bootstrap samples over tasks, retaining all available condition pairs within each sampled task. The 17 conditions of a task are not treated as independent observations. GPT-5.6 has a mean CQ change of +0.62 but an MAE of 16.26. DeepSeek-V4.1 RES rises from 1.54 to 16.62.

We also compare the sign of each single pressure effect relative to baseline across rounds, using tasks with all four required trajectories. Agreement across six pressures is below.

ModelCQ directions agreeRES directions agree
GPT-5.61/65/6
Sonnet-55/66/6
Gemini-3.85/66/6
DeepSeek-V4.13/61/6

Agreement means the same negative, zero, or positive direction.

For the main experiment, Appendix H.1 reports SD of pressure minus neutral CQ and RES differences and task bootstrap intervals. Each operational efficiency bootstrap sample recomputes the pooled counts and ratios. Main experiment uncertainty and changes across rounds are therefore reported separately.

Aggregate completion quality is stable across rounds for GPT-5.6, Sonnet-5, Gemini-3.8, and DeepSeek-V4.1, with changes of 0.62 to 2.98 points and every interval including zero, while Qwen3.8 shifts by -23.48 points. Sonnet-5 and Gemini-3.8 also reproduce 5 of 6 CQ directions and all 6 RES directions. Our robustness claims therefore rest on the aggregate scores of these four models, and Section 5.3 and Appendix L report trajectory variation and effect directions separately.

Q4

We agree that score changes need to be explained through specific behaviors. Appendix F separates CQ and RES into their components. Averaged over the six single pressures, the component means give these contributions to the CQ change, in points.

ModelTCCFCRPCRTotal
Gemini-3.8-0.30-1.35-0.63-2.28
GPT-5.6-1.40-0.10+0.23-1.27
Sonnet-5-1.98-1.77+1.08-2.67
Qwen3.8-0.87-2.28+0.37-2.77
DeepSeek-V4.1-2.69-1.05+0.26-3.48

Lost full completion drives the decline for Gemini-3.8 and Qwen3.8, lost end-state coverage drives it for GPT-5.6 and DeepSeek-V4.1, and Sonnet-5 loses both, while PCR rises for four models and offsets part of the loss. Unsafe successes fall for all five models, so RES declines through fewer unsafe actions while completion declines too.

The trajectory analysis connects these scores to care requirements. DeepSeek-V4.1 loses communication or ownership obligations in 20/31 eligible pairs in Appendix G. Under time pressure, Sonnet-5 loses required follow up in 5/15 pairs without unsafe execution in Table 5. Eligibility requires safely completed follow up at baseline, so this is a loss of an obligation the same model previously fulfilled.

Appendix K.2 provides a step by step pair in which the agent reads the records but repeatedly indexes, reopens, and rereads instead of drafting or publishing the reconciliation. The case stays open. Figure 7 further separates completion, process conformity, and unsafe execution. Together, the component scores and trajectories distinguish fewer unsafe actions from completed care.

Thank you for the detailed rebuttal and the additional experiments. The action-instruction control, the second evaluation round with uncertainty estimates, and the component and trajectory analyses address my questions. I will maintain my score.

Reviewer_PgoUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a benchmark for measuring "medical process integrity under workplace pressure". Medical agent safety is an important topic and this benchmark attempts to capture something that is not well explored - the effect of workplace pressure. The benchmark contains 200 medical tasks from fifive workflow functions and 17 experimental conditions. There are four main findings.

Critical strengths

The paper is creative. The idea of testing LLMs in this way is interesting (I have some concerns/questions that I will discuss below) but I do want to credit the authors for tackling an interesting problem and designing a benchmark around it.

The topic is important. The use of agents in clinical settings has a lot of potential but there are many important considerations when deploying such a system in a high-stakes, safety critical setting.

The paper is generally well motivated and presented. However, I did feel that the beginning of the paper (introduction and design) was better presented than the results.

The figures and tables are not overly complex.

Critical weaknesses

While the paper is interesting, it was surprising to me how difficult it was to understand exactly how the workplace pressure was actually implemented in the context of the experiments. As far as I can understand the pressure is represented as prompt language - which feels artificial - why does it need to represented like that even if the situation is urgent, for example. "The ordered conditions also use both scheduled pressure messages and order expressed in prompt text." How do the authors argue that this is a realistic representation of pressure? And why to AI systems need to presented with that "pressure" when their sense of time is very different from humans?

I find the cartoony presentation of the results (e.g. Fig 5) a bit distracting and not helpful. I would prefer to see tables with more results and more robust error/variance metrics. I find Tables 2-4 difficult to intepret given that the conditions are only represented as letters and the metrics are not defined well in the captions.

The authors state several times that " Physicians reviewed all tasks for medical plausibility and consistency", this is good and any human annotation increases the value of a benchmark. However, there is very little detail about the nature of the reviews - who the physicians were that did them and what they did.

I would argue some important foundational related work is missing. I would encourage the authors to be a bit more thorough in their background.

Questions for authors

The captions of the figures and tables are very limited and make it harder for the reader to interpret them, can you make them clearer?

Can you provide more justification for the framing of the problem and why prompting an LLM with time pressure is a valid thing to do and reflects how it would be used in real life?

Can you provide more details about the nature of the physician annotations - what profile did the physicians have, how many of them were there, how many years of experience did they have, how were they compensated, how many physicians reviewed each case, etc. many more details are needed.

2 comments

Thank you for recognizing the importance of the problem. We added repeated evaluations and uncertainty estimates, improved figures, and clarified clinical review. We label your four weaknesses W1 to W4.

W1 & Q2

We ask whether a medical agent preserves safe and complete care after receiving workplace urgency information. This does not require an LLM to experience time or psychological stress as a person does. We examine whether the context changes evidence checks, action selection, and completion of downstream obligations. The test concerns its response to communicated urgency, without reducing the actual time or decision budget.

Section 3.3 and Appendix D.1 give the clinical basis and example texts, and physicians rated the six contexts at 3.49 to 4.42 on a 1–5 realism scale in Section 4. Section 3.3 now states delivery through pressure_context at each decision, scheduled in Appendix D.4. Evidence, authorization, tools, completion requirements, and decision budget remain fixed. Appendix D.3 separates scheduled messages from order stated in the prompt.

The Introduction adds medical agent examples with resource, time, and budget constraints. Section 4 and Appendix M add 300 action instruction runs to compare organizational context with explicit guidance.

W2 & Q1

We agree that readers need interpretable numerical results and uncertainty estimates. Appendix H tabulates every paired change behind Figures 5 and 6, with SD and task bootstrap 95% intervals in H.1. Figure 7 separates completion, unsafe execution, and process conformity. Table 2 reports the baseline components, while Tables 3 and 5 give obligation losses and follow-up outcomes with explicit denominators. The captions of Tables 2 and 3 and Figures 5 and 6 now define labels, units, and eligibility.

We also added separate statistical analyses for the main evaluation and the additional round. Section 5.1 and Appendix H.1 report standard deviations of paired CQ and RES changes and task bootstrap 95% intervals for mean changes. For operational efficiency, each bootstrap sample recomputes pooled counts and ratios.

We also ran a second round on 20 tasks covering all 17 conditions. Section 5.3 (Table 4) and Appendix L report additional minus original CQ below. SD and MAE describe paired differences, and CI is the task bootstrap 95% interval for the mean.

ModelPairsMeanSDMAE95% CI
GPT-5.6340+0.6222.0316.26[-2.41, +3.42]
Sonnet-5338+0.6819.8210.49[-1.20, +2.76]
Gemini-3.8339+2.9817.3112.07[-0.75, +6.46]
DeepSeek-V4.1340+2.1319.5515.57[-1.66, +5.85]
Qwen3.8339-23.4828.0629.58[-32.01, -15.13]

Resampling retains all conditions within a task. CQ is stable across rounds for four models, with every interval including zero, while Qwen3.8 shifts by -23.48 points.

W3 & Q3

Clinical review involved three physicians, each with at least three years of clinical experience. Each of the 200 cases was reviewed by two physicians, about 130 to 140 cases per physician. Disagreements went to the third physician and were resolved through discussion. The two initial reviews agreed on 157 of 200 cases, and 27 cases were revised. Physicians were compensated at an hourly rate for their review time. Because payment was in local currency, we report it relative to salary, and the three physicians' total compensation corresponded to 24.8%, 30.5%, and 22.0% of their respective monthly salaries.

The review addressed the task specifications and care requirements. As described in Section 4 and Appendix C.4, physicians assessed the plausibility of the evidence and patient state, whether authorization requirements and the expected action or restraint fit the case, and whether downstream obligations matched the care process. This included monitoring, review, communication, handoff, documentation, and follow-up where applicable. The Ethics Statement explains its role in the specifications and criteria.

This task review is distinct from scoring model outputs. Completion and safety scores are computed from recorded tool events and final workspace states using deterministic rules in Appendix F.

W4

We have expanded the clinical background and clarified its connection to agent evaluation. Section 3.2 connects the workflow functions to prior work on medical workflows, evidence qualification, record reconciliation, and responsibility transfer. Section 3.3 links pressure categories to healthcare studies. The new Introduction examples connect time and resource requirements directly to medical agent research (Yuan et al., Meincke et al., and Nori et al.). Section 2 places our evaluation alongside work on medical agents, unfinished care, and agent safety under workplace pressure, and now adds evidence that errors concentrate at care transitions.

Reviewer_PgoUOpenReview ↗

Response to authors

I would like to thank the authors for their response. They have addressed most of questions in some form. I do concur with the other reviewer that there are concerns about the ecological validity of the data (scenarios and tasks), this was the basis for the first weakness I highlighted. While the authors have allayed this concern somewhat I don't think they have fully addressed then to convince me that these are the strongest examples. On balance I still think the paper isn't quite above the bar for the conference but I want to give the authors credit for the response. It is a tough call, but I think my score is still representative of the overall judgement.

Reviewer_eAtCOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces MedicalPressureBench, which includes 200 medical tasks, and then tests how six different kinds of "work place pressures" impact the performance of agentic systems. They claim that while single pressures decrease performance, combinations of pressures can have weakened effects. They claim that this benchmark can be used as a safety assessment of how pressures impact agent systems.

Critical strengths

-Critical strengths include writing 200 novel cases for assessment (which reduces the chance of the testing data being in model training data). -LLMs were not used for judging, which can be a point of bias; this is a critical strength in design choice made by the authors.

Critical weaknesses

The biggest weakness is ecological validity here. I'm not sure that any of this actually represents how real world medical workflows occur (note that I know the healthcare system incredibly well, both as a practitioner as someone who has built implemented AI tools). For example: -Very little is said about what the 200 cases are. They do not use real patient data, and there is no way to know if the synthetic scenarios are actually medically valid -the five medical workflow functions do not appear to based on any prior work assessing medical workflows. The authors do not justify why they picked these particular workflows and whether they have any real validity in medical practice. -There is very little said about how the six different pressures are presented or injected, whether variation in the severity of a single pressure has an impact. -It seems odd to me that multiple different kinds of pressure may negate one another, but very little is said about why this might be true. It again makes me question the validity of the entire set up.

Questions for authors

How did you determine the 5 medical workflows? How did you validate that your 200 cases actually reflect medical practice? How did you represent the pressures? How do you know they are realistic for what happens in medical practice?

2 comments

Thank you for examining the benchmark from your clinical and AI implementation experience, and for recognizing the newly written cases and the evaluation without an LLM judge. We agree that its value depends on a clear medical basis for the tasks and evaluation requirements. We address case validity and workflow selection here, with pressure representation and combination results in a separate comment.

W1 & Q2

Thank you for asking how the cases were validated. All 200 tasks are synthetic, and no real patient records were used. We asked physicians to review the medical information, clinical state, and care requirements of each synthetic case, beyond its narrative. The procedure is below.

Review itemProcedure
Physicians3, each with at least 3 years of clinical experience
Initial review2 physicians per case across all 200 cases
DisagreementThird physician review followed by discussion
Initial agreement157 of 200 cases (78.5%)
Revised after review27 cases
Medical contentPlausibility of the medical information and clinical state specified in each synthetic case
Action boundaryAuthorization and expected action or restraint
CompletionRequired monitoring, review, communication, handoff, records, and follow up

These checks connect the case to its expected endpoint. A task can require an authorized action followed by monitoring, or require withholding release while arranging review. The review therefore concerns both whether the immediate decision fits the evidence and whether the remaining obligations fit the care process. An acceptable final answer alone does not establish completion. Tool events and final states determine whether the specified requirements were fulfilled, using deterministic rules with no LLM judge.

Construction starts from a workflow failure, identifies the obligation that prevents or resolves it, and maps that obligation to an observable workspace state or tool event. We also checked that each synthetic task retains the same records, tools, completion requirements, and action guard across pressure conditions. Clinical review addresses the medical content, while the invariance audit addresses whether comparisons retain the same task. Section 4 and Appendix C document these checks.

We use synthetic cases to hold evidence and appropriate endpoints fixed while varying workplace context. Each pressure effect is therefore a matched comparison on the same physician-reviewed task. We position the benchmark as a controlled medical workflow evaluation, which measures how agents respond to workplace context under clinically reviewed requirements.

W2 & Q1

We agree that these clinical distinctions need a clear basis. The five functions separate different medical transitions, independently of disease or specialty. The table explains the relevant evidence and links directly to each study.

FunctionClinical distinction and supporting evidence
ExecuteCompleting an action differs from retrieving information. MedAgentBench, Jiang et al., NEJM AI 2025 includes test and medication ordering tasks.
GateDeciding whether to proceed differs from executing an action. Guideline model comparison, Peleg et al., JAMIA 2003 distinguishes decision and action models.
QualifyA result must be assessed before subsequent use. Positive blood cultures, Weinstein et al., 1983 distinguishes true bloodstream infection from contamination.
ReconcileConflicting records require resolution before use. Medication reconciliation trial, Schnipper et al., 2009 examines medication discrepancies and potential adverse drug events.
HandoffInformation must reach the clinician taking over care. Handoff program, Starmer et al., NEJM 2014 evaluates standardized oral and written handoffs and associated medical errors.

These studies ground each distinction, and we translate them into observable completion requirements. Execute includes monitoring and a record after action. Gate includes a review path after a hold. Qualify includes a calibrated status and repeat plan. Handoff requires a named recipient who takes ownership. Reconcile requires one current version before further action or release. Figure 3 and Appendix B also show these requirements.

The 40 tasks per function give balanced coverage, so failures can be compared across functions and located at the transition where a workflow breaks.

Please see the other comment for the remaining issues.

This comment addresses pressure representation, weaker combination effects, and clinical relevance.

W3 & Q3

We agree that pressure delivery and severity need to be distinguished. As Section 3.3 now states, pressure is organizational context in pressure_context at each decision. For example, the time context describes a narrowing interval while the surrounding schedule continues. Neutral, single pressure, and simultaneous conditions repeat the same context from decision 1. For archived order conditions with scheduled messages, the first pressure appears at decisions 1 to 5 and the second at 6 to 10. Reversing the condition exchanges them.

The workplace demand changes, while synthetic case information, authorization, tools, completion requirements, and the ten decision budget stay fixed. We do not shorten actual execution time or require an LLM to experience human psychological stress. We examine whether this information changes evidence checks, action selection, and completion of downstream obligations.

On severity, we have not calibrated the six categories to a common scale or run a single pressure intensity gradient. The results describe behavior under specified contexts, not an effect per intensity level. Time and performance combine urgency with throughput demands, while regulatory and peer contexts combine scrutiny with team norms. We do not assume a combination is stronger than a single pressure. The instruction control in Section 4 (Table 1) compares context with explicit action requirements.

W4

Thank you for asking why combinations can have weaker effects. Different demands can change which action an agent takes or where it stops, and their number is not a common intensity scale.

We compared tools and final states for the same model and synthetic task under two single pressures and their combination, with records and action requirements fixed.

Paired taskUnder each single pressureUnder the combination
Gemini-3.8, pathology handoff, regulatory and peerCalls finish with obligations missing, completing 5 of 7 requirementsNo finish call, but still only 5 requirements completed, with no surgeon message or case entry
Gemini-3.8, prenatal qualification, regulatory and peerRecords discrepancies without publishing statusCalls the publication tool, crossing the configured boundary and recording unsafe execution
DeepSeek-V4.1, preparation release, time and performanceCalls the release tool and records unsafe executionRecords a hold without calling release, avoiding that unsafe execution but leaving requirements incomplete

The first pair reduces premature finish violations without restoring missing care. The second moves from recording a problem to publication and increases unsafe execution. The third reduces recorded risk by withholding release, while remaining incomplete.

The changes also differ across tasks. Among 200 common GPT-5.6 tasks, combined completion quality exceeds both single pressures in 32 tasks and falls below both in 42. An aggregate mean therefore combines opposing task-level changes.

These trajectories show that a weaker combined effect comes from redirected workflows, each leaving its own gap in completion or safety. Obs. 4 in Section 5.2 now presents the three pairs, with the full comparison in Appendix I.1.

Q4

Thank you for asking about clinical relevance. The three physicians also rated each pressure context for realism on a 1–5 scale, as Section 4 now reports.

PressureMeanSD
Regulatory4.110.44
Time4.420.31
Authority3.840.64
Peer3.610.72
Precedent3.490.56
Performance4.250.40

The categories draw on healthcare studies of scrutiny, hierarchy, norms, and preceding cases. Medical agent research provides direct examples of time, resource, and cost demands.

Asclepius, Findings of EMNLP 2026 requires agents to manage arriving emergency patients with limited beds and staff, introducing time and resource demands across patients.

Clinical simulation, Meincke et al., 2026 requires an autonomous agent to assess and treat patients within 20 minutes.

Sequential Diagnosis with Language Models, 2025 includes diagnostic test costs in MAI-DxO decisions, introducing budget and efficiency requirements alongside accuracy.

These studies show that time, resource, and cost demands shape medical agent tasks. Our claims concern how agents respond to such organizational demands when the medical task stays fixed, which the matched design measures directly.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.