One Agent
or Many?

AI systems that conduct structured patient interviews and draft diagnostic interpretations for clinician review are moving from research prototypes toward everyday practice. A central design question remains unsettled: should such a system be built as one comprehensive agent, or decomposed into several specialized agents that hand work to one another?

Multi-agent architectures promise role specialization and tighter control. They also add complexity, latency and cost, and create a new opportunity for information loss at every handoff. This white paper summarizes a controlled pilot study, conducted at Maastricht University during an internship with Ozyra, that put the question to the test under unusually clean conditions: both architectures used the same foundation model, the same retrieval corpus of Dutch general-practice (NHG) guidelines, the same standardized synthetic patients and the same blinded evaluation rubric. Only the control structure differed.

The results challenge a common assumption. Multi-agent decomposition did not produce a statistically significant improvement in average report quality or in clinical safety. Where the architectures parted ways was reliability: multi-agent outputs were markedly more consistent when the same clinical case was run repeatedly, while the single agent alternated between excellent and mediocre reports for identical presentations.

For healthcare organizations evaluating clinical AI, this reframes the architecture question. The choice may matter less for how good a system is on average — and more for how predictably it behaves. Predictability is what determines whether a system can be validated, audited and safely embedded in a clinician-in-the-loop workflow.

Large language models can now conduct a structured anamnesis: receive a patient’s opening complaint, ask guideline-derived follow-up questions, check red flags, and draft a diagnostic interpretation for a clinician to review. The clinical value proposition is time — inquiry and documentation prepared before the consultation — while the clinician retains every decision.

Behind such systems sit two competing design philosophies. A single-agent system (SAS) gives one comprehensive model instruction responsibility for the entire workflow: identifying the relevant guideline, interviewing the patient, screening for red flags, and writing the interpretation. A multi-agent system (MAS) decomposes the same workflow into specialized roles — here, a guideline matcher, an anamnesis interviewer and an interpretation writer — that pass structured information from one to the next.

Multi-agent designs were originally motivated by real weaknesses of earlier models, which used long contexts unreliably and struggled with complex, multi-constraint instructions. But foundation models have improved rapidly, and recent research questions whether multi-agent advantages always reflect genuine architectural benefit rather than simply additional computation. Meanwhile, every added agent brings engineering complexity, latency, cost, and a new place for clinical information to be lost in a handoff.

For anyone building or procuring clinical AI, the question is therefore empirical rather than philosophical: with today’s strong models, does decomposition still buy anything — and if so, what exactly?

Synthetic patients were used deliberately: real consultations involve sensitive health data, and standardized cases allow a controlled, repeatable comparison in which architecture is the only variable. The trade-off — reduced ecological validity — is addressed under Limitations.

3.1Average quality: high in both — and not the differentiator

The clinician rated reports from both architectures near the top of the scale: a mean total score of 16.15 of 18 for the multi-agent system against 15.15 for the single agent. The trend favored decomposition, and the LLM judge pointed the same way, but neither difference was statistically significant (clinician p = 0.172). Read plainly: with a strong foundation model and careful prompt engineering, a single agent already produces guideline-grounded reports of high average quality. Decomposition is not a prerequisite for a good clinical report.

3.2Clinical safety: comparable

Safety was scored as its own rubric dimension so that risk handling could not disappear inside an aggregate. Mean safety scores were 2.60 (multi-agent) versus 2.50 (single agent) of 3, with no significant difference and the same picture from the LLM judge. Both systems were built to interview against guideline red flags and to route every interpretation to a clinician — and neither showed a systematic safety edge over the other.

3.3Consistency: where the architectures part ways

Each architecture ran every case twice, which allows a simple but clinically meaningful question: how far apart do two runs of the same system land on the same patient? For the single agent, the mean within-case gap was 3.10 points on the 18-point scale — in the widest instance, the same presentation yielded a near-perfect report on one run and a clearly deficient one on the next. The multi-agent system’s mean gap was 0.70 points.

This difference reached conventional significance in the exploratory analysis (p = 0.047), and the independent LLM judge found the same pattern even more sharply (0.73 vs. 3.43 points, p = 0.006), agreeing with the clinician’s direction in six of seven non-tied topics. A plausible mechanism: structured handoffs act as stabilizers. Each multi-agent stage receives a constrained, machine-readable state — a matched guideline, a confirmed patient summary — rather than an open-ended conversation, which narrows the space of trajectories the system can take. Given the pilot scale, this finding is hypothesis-generating rather than proven. But it is the study’s clearest and most coherent signal.

TABLE 1  —  Means over ten topic-level values; p-values are unadjusted exploratory Wilcoxon signed-rank tests. Directional agreement between LLM judge and clinician: 7/7 topics for overall quality, 6/7 for consistency, 3/4 for clinical safety.

This was a pilot: ten clinical topics, forty reports, one clinician rater, one foundation model, and synthetic rather than real patients. The statistics are exploratory, without correction for multiple testing, and the consistency finding — however coherent across two independent evaluators — requires confirmation at larger scale, with multiple independent clinicians, independently initialized patient simulations, and additional model families. Real patients are also less tidy than simulated ones: incomplete, ambiguous, sometimes contradictory, often worried. None of the findings here should be read as proof that one architecture is clinically superior. They should be read as a disciplined first measurement of where the differences actually lie — and, just as importantly, of where they do not.

Ozyra builds guideline-grounded digital anamnesis, monitoring and coaching for healthcare providers, with clinician review at the core of every workflow. Three engineering commitments follow directly from this research.

Reliability-first architecture. We use decomposition where it demonstrably stabilizes output and simplicity where it does not, rather than treating agent count as a feature. Architecture is a means; predictable clinical behavior is the requirement.

Auditable handoffs. Where workflows are decomposed, state passes between stages as structured, inspectable data — the property this study associates with consistency, and the one that makes system behavior traceable after the fact. In regulated medical software, traceability is not optional.

Evaluation as infrastructure. Blinded clinician scoring, separated quality, safety and consistency outcomes, and repeated-run consistency tracked as a release criterion — built into development, not bolted on afterwards.

We publish our results, including the null results, because clinical AI should be chosen on evidence — not on architecture diagrams.

This white paper is based on the Bachelor’s thesis “Is Multi-Agent Decomposition Necessary? A Comparative Study of Single-Agent and Multi-Agent LLM Architectures for Guideline-Grounded Clinical Report Generation” by Zhengzhong Huang, Department of Advanced Computing Sciences, Faculty of Science and Engineering, Maastricht University (supervisors: Max Sondag and Jerry Spanakis), written during an internship with Ozyra, 2026. The full thesis — including complete system prompts, the evaluation protocol and all statistical analyses — is available on request.