Multi-Agent Clinical Diagnosis (MACD)

• MACD is a clinical decision-support framework that distributes diagnostic reasoning across multiple specialized AI agents for evidence-based consensus.
• It employs explicit role specialization, structured communication, and sequential information gathering to mimic multidisciplinary team reviews and iterative bedside reasoning.
• MACD systems improve diagnostic accuracy and interpretability by integrating protocol-driven collaboration, memory-backed hypothesis verification, and privacy-preserving data sharing.

Searching arXiv for recent MACD-related papers to ground the article with relevant citations. Multi-Agent Clinical Diagnosis (MACD) denotes a class of clinical decision-support frameworks in which multiple specialized agents—typically LLM-based, but sometimes coupled to retrieval, simulation, or multimodal modules—jointly perform diagnostic reasoning, information gathering, evidence interpretation, and consensus formation. In the current literature, MACD systems are characterized by explicit role decomposition, structured inter-agent communication, and diagnostic workflows that more closely resemble consultation, multidisciplinary team review, or iterative bedside reasoning than single-shot prompting. Across recent work, MACD has been instantiated as hierarchical GP–specialist systems, debate-based committees, dynamic interactive diagnostic loops, self-evolving inquiry agents, and privacy-preserving cross-institutional collaborations (Zhan et al., 1 Mar 2026, Sanghvi et al., 2 Jun 2026, Ge et al., 27 Mar 2026, Wang et al., 11 Jun 2026).

1. Conceptual scope and defining characteristics

The core premise of MACD is that clinical diagnosis is not a single inference from a static prompt, but a composite process involving hypothesis generation, uncertainty management, information acquisition, evidence verification, and differential adjudication. Several recent systems explicitly reject the “one-shot” formulation and instead model diagnosis as a multi-turn or multi-stage procedure in which agents ask questions, order tests, examine evidence, revise beliefs, and only then produce a final diagnosis (Shang et al., 3 Jul 2025, He et al., 30 Jan 2026, Sanghvi et al., 2 Jun 2026).

A common architectural pattern is role specialization. MedCollab partitions the full agent pool $A$ into clinical-specialist agents and examination agents, with $A=\mathrm{Clin}\cup \mathrm{Ex}$ and $\mathrm{Clin}\cap \mathrm{Ex}=\emptyset$, then dynamically recruits a case-specific subset using a recruitment function $R:S\mapsto 2^A$ driven by $P(a_i\mid S)$ (Zhan et al., 1 Mar 2026). MDAgents similarly distinguishes a Moderator, Recruiter, Expert Clinician Agents, and a Final Decision-Maker, with the collaboration structure varying across low-, moderate-, and high-complexity tasks (Kim et al., 2024). ClinicalAgents uses an Orchestrator, Hypothesis Generator, Evidence Checker, Backtracker, and additional domain experts, embedding diagnosis in an MCTS-controlled hypothetico-deductive loop supported by dual memory (Ge et al., 27 Mar 2026).

A second defining feature is explicit coordination rather than implicit ensemble averaging. MACD systems often include supervisor, manager, judge, evaluator, or orchestrator modules that route cases, regulate dialogue, trigger reassignment, or arbitrate disagreement. Mixed-vendor Multi-Agent Conversation (MAC) employs a fixed supervisor overseeing up to three doctor agents in turn-taking dialogue until a Top-$K$ differential diagnosis list stabilizes or a turn limit is reached (Yuan et al., 14 Feb 2026). SOAP-note problem detection uses a Manager agent that dynamically assigns specialists and monitors iterative debate rounds until consensus or fallback aggregation (Lee et al., 29 Aug 2025). CAMP adds a hybrid router that sends candidate diagnoses through strong-consensus, fallback, or evidence-based arbitration paths (Lu et al., 31 Mar 2026).

A third characteristic is interpretable intermediate structure. Rather than emitting only final labels, many MACD systems expose rationales, evidence spans, support scores, or causal structures. MedCollab adopts an IBIS argumentation protocol in which agents produce tuples $(I,P_k,\mathrm{Arg}_k,E_k)$, with a traceability constraint enforcing that each cited evidence item supports the corresponding diagnostic position via $\phi:\mathcal{E}\to P$ (Zhan et al., 1 Mar 2026). Counterfactual multi-agent reasoning quantifies evidence support with the Counterfactual Probability Gap, making explicit how diagnosis confidence changes when findings are negated, removed, weakened, or replaced (You et al., 29 Mar 2026). MDTRoom externalizes evidence provenance, hypothesis evolution, and conflict structure into a structured workspace rather than linear transcript logs (Kuai et al., 30 Mar 2026).

These recurring design choices suggest that MACD is best understood not merely as “multiple LLMs answering the same question,” but as an engineering paradigm for distributing clinical reasoning across differentiated roles, explicit protocols, and inspectable deliberation states.

2. Agent architectures and organizational patterns

Recent MACD systems differ substantially in how they decompose labor across agents, but the major patterns are hierarchical triage, panel deliberation, sequential environment interaction, and modular pipeline decomposition.

Hierarchical triage architectures mirror hospital consultation pathways. KG4Diagnosis uses a two-tier structure $A=\{\mathrm{GP},S_1,\dots,S_n\}$, in which a GP agent performs initial assessment and triage, referring to specialist agents when $P_{\mathrm{GP}}(x\mid q)<\tau$ or when a condition lies in a specialist-only set (Zuo et al., 2024). MedCollab extends this idea with dynamic recruitment over both specialist and examination agents, allowing the GP to recruit all agents with $A=\mathrm{Clin}\cup \mathrm{Ex}$0 (Zhan et al., 1 Mar 2026). MDAgents adds complexity-sensitive routing, using solo, MDT, or ICT collaboration depending on whether the moderator classifies the case as low, moderate, or high complexity (Kim et al., 2024).

Panel deliberation architectures assemble a consultation team and resolve disagreement through voting, debate, or arbitration. Mixed-vendor MAC uses three doctor agents instantiated with o4-mini, Gemini-2.5-Pro, and Claude-4.5-Sonnet under a fixed supervisor (Yuan et al., 14 Feb 2026). CAMP’s attending physician dynamically assembles a specialist panel $A=\mathrm{Clin}\cup \mathrm{Ex}$1, then each specialist votes $A=\mathrm{Clin}\cup \mathrm{Ex}$2, $A=\mathrm{Clin}\cup \mathrm{Ex}$3, or $A=\mathrm{Clin}\cup \mathrm{Ex}$4 on each candidate diagnosis, after which a hybrid router determines whether to accept by strong consensus, defer to the attending physician, or launch evidence-based arbitration (Lu et al., 31 Mar 2026). In SOAP-note problem detection, five specialists debate binary target-problem presence with thresholded consensus $A=\mathrm{Clin}\cup \mathrm{Ex}$5, team reassignment, and managerial fallback (Lee et al., 29 Aug 2025).

Interactive diagnostic-loop architectures model diagnosis as sequential evidence collection under partial observability. DynamiCare formulates MACD as a sequential, partially observable multi-agent decision process with latent patient state $A=\mathrm{Clin}\cup \mathrm{Ex}$6, active team $A=\mathrm{Clin}\cup \mathrm{Ex}$7, visit log $A=\mathrm{Clin}\cup \mathrm{Ex}$8, and cumulative reward balancing diagnostic accuracy and interaction cost (Shang et al., 3 Jul 2025). EvoClinician uses a Patient agent and Examination agent as environment “gatekeepers,” while an Actor agent issues one JSON action per turn from $A=\mathrm{Clin}\cup \mathrm{Ex}$9, and separate Process Grader and Evolver agents adapt strategy at test time (He et al., 30 Jan 2026). MedAgentSim uses doctor, patient, and measurement agents in a turn-based stochastic game, with action spaces $\mathrm{Clin}\cap \mathrm{Ex}=\emptyset$0, $\mathrm{Clin}\cap \mathrm{Ex}=\emptyset$1, and $\mathrm{Clin}\cap \mathrm{Ex}=\emptyset$2 for the doctor (Almansoori et al., 28 Mar 2025).

Pipeline-decomposition architectures split diagnosis into upstream perception or retrieval stages and downstream clinical reasoning. LungNoduleAgent comprises a Nodule Spotter, a Simulated Radiologist, and a Doctor Agent System for malignancy reasoning using images, masks, CT reports, and a pathology knowledge base (Yang et al., 26 Nov 2025). The secondary-headache system routes free-text vignettes to seven red-flag specialist agents—such as thunderclap, meningismus, papilledema, and focal deficits—after orchestrator feature extraction (Wu et al., 3 Dec 2025). MeDxAgent decomposes interactive consultation into Question Agent, Summarizer, Diagnosis Agent, Evidence-Gap Agent, Patient Simulator, and Judge (Sanghvi et al., 2 Jun 2026).

These heterogeneous designs indicate that MACD is an umbrella category rather than a single algorithm. What unifies them is not a fixed topology, but the use of multiple agents as explicit computational counterparts to triage, consultation, testing, synthesis, and review.

3. Reasoning protocols, evidence handling, and consensus mechanisms

A central research question in MACD concerns how agents should communicate and how their outputs should be combined. The literature has moved beyond unstructured conversation toward formalized evidence grounding, weighted voting, audit rules, and counterfactual verification.

MedCollab provides one of the most explicit formalizations. Within its IBIS-Structured Argumentation Protocol, each agent $\mathrm{Clin}\cap \mathrm{Ex}=\emptyset$3 produces IBIS tuples $\mathrm{Clin}\cap \mathrm{Ex}=\emptyset$4, where $\mathrm{Clin}\cap \mathrm{Ex}=\emptyset$5 is a diagnostic hypothesis and $\mathrm{Clin}\cap \mathrm{Ex}=\emptyset$6 is a supporting evidence subset. The traceability constraint requires that for every cited evidence item $\mathrm{Clin}\cap \mathrm{Ex}=\emptyset$7, $\mathrm{Clin}\cap \mathrm{Ex}=\emptyset$8, ensuring that every proposed position is grounded only in explicitly cited evidence (Zhan et al., 1 Mar 2026). MedCollab then constructs a Hierarchical Disease Causal Chain $\mathrm{Clin}\cap \mathrm{Ex}=\emptyset$9, a directed acyclic graph over validated positions, with optional logical operators $R:S\mapsto 2^A$0 and probabilistic annotations $R:S\mapsto 2^A$1 on edges (Zhan et al., 1 Mar 2026).

Consensus mechanisms vary widely. MedCollab defines a positive weight function $R:S\mapsto 2^A$2, binary support indicators $R:S\mapsto 2^A$3, and a consensus score $R:S\mapsto 2^A$4, accepting a position only if $R:S\mapsto 2^A$5. It further audits arguments against a rule set $R:S\mapsto 2^A$6, computes per-agent inconsistency scores $R:S\mapsto 2^A$7, and updates weights across rounds via $R:S\mapsto 2^A$8 until accepted positions stabilize (Zhan et al., 1 Mar 2026). This is notable because consensus is not purely a vote count, but is shaped by logic auditing and reliability penalties.

CAMP replaces binary voting with principled abstention. For each candidate diagnosis $R:S\mapsto 2^A$9, specialists produce counts $P(a_i\mid S)$0, $P(a_i\mid S)$1, and $P(a_i\mid S)$2 for Keep, Refuse, and Neutral votes, respectively, and the router uses these counts to select among strong consensus, weak-consensus fallback, or attending-physician arbitration (Lu et al., 31 Mar 2026). The arbitration stage weighs argument quality via $P(a_i\mid S)$3 and $P(a_i\mid S)$4 rather than merely tallying votes (Lu et al., 31 Mar 2026). This design directly addresses a common weakness of flat majority voting: specialists outside their competence can abstain rather than distort the outcome.

Counterfactual multi-agent reasoning introduces a different verification primitive. Specialists generate counterfactual case edits—Negate, Remove, Replace, Weaken, Intensify, Insert—and evaluate diagnosis support using the Counterfactual Probability Gap

$P(a_i\mid S)$5

High CPG indicates that a modified finding was critical to supporting diagnosis $P(a_i\mid S)$6; low CPG indicates a non-discriminating or incidental feature (You et al., 29 Mar 2026). These counterfactual signals guide multi-round discussion, allowing agents to challenge unsupported hypotheses with explicit evidence-perturbation tests.

Other frameworks emphasize compatibility of information flow. The Optimization Paradox study decomposes diagnosis into information gathering $P(a_i\mid S)$7, interpretation $P(a_i\mid S)$8, and differential diagnosis $P(a_i\mid S)$9, and shows that a Best-of-Breed system with stronger component-level metrics nonetheless underperforms diagnostically because of a large compatibility gap $K$0 arising from heterogeneous communication styles and unexpected interactions (Bedi et al., 6 Jun 2025). This is a reminder that improved sub-agents do not guarantee improved MACD if interfaces are poorly aligned.

A plausible implication is that MACD performance depends at least as much on protocol design as on raw base-model strength. Formal grounding, abstention, auditability, and interface compatibility appear repeatedly as mechanisms for converting multiple agents into a coherent diagnostic system rather than a noisy ensemble.

4. Dynamic, interactive, and sequential diagnosis

One major strand of MACD research treats diagnosis as an online information-gathering problem rather than a deliberation problem over fixed records. In these systems, agents must decide what to ask, what to test, when to stop, and how to balance information gain against cost.

DynamiCare formulates MACD as a Dec-POMDP-like process with latent patient health state, observable initial record $K$1, round-specific team $K$2, action space $K$3 consisting of natural-language queries or diagnoses, observation space $K$4, and an objective

$K$5

The Central Agent updates team composition dynamically based on the visit log, specialists propose either next questions or diagnoses with confidences, and the patient system maps queries to JSON-aligned responses derived from MIMIC-Patient records (Shang et al., 3 Jul 2025). In open-ended diagnosis on 500 MIMIC-Patient cases, multi-agent GPT-4.1 achieved Hit@5 $K$6, Hit@10 $K$7, Rec@5 $K$8, Rec@10 $K$9, with Ave-Q $(I,P_k,\mathrm{Arg}_k,E_k)$0, outperforming the single-agent GPT-4.1 baseline across all reported metrics (Shang et al., 3 Jul 2025).

MeDxAgent studies prompt-, flow-, and agent-level choices for interactive diagnosis on MeDxBench, a 4,421-case benchmark across 20 specialties. Its high-level consultation flow enforces a demographics-first question at turn 1, uses evidence-guided QA before turn 10, evidence-guided differential QA after turn 10, updates a summary every turn, and stops early if any diagnostic confidence reaches $(I,P_k,\mathrm{Arg}_k,E_k)$1 (Sanghvi et al., 2 Jun 2026). Its reported accuracy is $(I,P_k,\mathrm{Arg}_k,E_k)$2, compared with a baseline of $(I,P_k,\mathrm{Arg}_k,E_k)$3, for a $(I,P_k,\mathrm{Arg}_k,E_k)$4 gain and $(I,P_k,\mathrm{Arg}_k,E_k)$5 oracle-gap closure (Sanghvi et al., 2 Jun 2026). The paper also reports that differential questioning at turn 2 collapses accuracy to $(I,P_k,\mathrm{Arg}_k,E_k)$6, while delaying it to turn 10 yields the best trade-off (Sanghvi et al., 2 Jun 2026), underscoring the importance of information-ordering policies.

EvoClinician frames multi-turn diagnosis as a POMDP with sparse end-of-episode judge score $(I,P_k,\mathrm{Arg}_k,E_k)$7, action costs $(I,P_k,\mathrm{Arg}_k,E_k)$8, and objective $(I,P_k,\mathrm{Arg}_k,E_k)$9 (He et al., 30 Jan 2026). Its “Diagnose–Grade–Evolve” loop gives per-action feedback labels such as HIGH_YIELD, LOW_YIELD, INEFFICIENT, and CRITICAL_ERROR, and uses prompt and memory evolution to improve future behavior without gradient-based RL (He et al., 30 Jan 2026). On Med-Inquire’s 915 NEJM/Lancet cases with gemini-3-pro, EvoClinician reports $\phi:\mathcal{E}\to P$0, $\phi:\mathcal{E}\to P$1, and $\phi:\mathcal{E}\to P$2, versus Static Prompt $\phi:\mathcal{E}\to P$3, $\phi:\mathcal{E}\to P$4, and $\phi:\mathcal{E}\to P$5 (He et al., 30 Jan 2026).

A related RL-oriented formulation appears in the consultation-flow framework using Doctor, Patient, and Examiner agents. The doctor selects actions from a hierarchical set spanning Inquiry, Examination, and Diagnosis phases, receives a small negative step cost and positive terminal reward for correct diagnosis, and is evaluated on the MVME benchmark (Wang et al., 19 Mar 2025). Full three-phase RL improves Qwen-Max symptoms score from $\phi:\mathcal{E}\to P$6 to $\phi:\mathcal{E}\to P$7, medical examinations from $\phi:\mathcal{E}\to P$8 to $\phi:\mathcal{E}\to P$9, diagnostic results from $A=\{\mathrm{GP},S_1,\dots,S_n\}$0 to $A=\{\mathrm{GP},S_1,\dots,S_n\}$1, diagnostic rationales from $A=\{\mathrm{GP},S_1,\dots,S_n\}$2 to $A=\{\mathrm{GP},S_1,\dots,S_n\}$3, treatment plan from $A=\{\mathrm{GP},S_1,\dots,S_n\}$4 to $A=\{\mathrm{GP},S_1,\dots,S_n\}$5, and entity-level F1 from $A=\{\mathrm{GP},S_1,\dots,S_n\}$6 to $A=\{\mathrm{GP},S_1,\dots,S_n\}$7 (Wang et al., 19 Mar 2025).

This interactive line of work broadens MACD beyond transcript debate. It situates diagnosis in sequential control, where agent collaboration is used not only to aggregate opinions but to decide what evidence should exist in the first place.

5. Knowledge augmentation, memory, and causal structuring

Another major theme in MACD is the introduction of structured knowledge and persistent memory to stabilize reasoning, reduce hallucination, and support reusable clinical expertise.

ClinicalAgents exemplifies memory-centric MACD. Its Working Memory $A=\{\mathrm{GP},S_1,\dots,S_n\}$8 stores evolving evidence, hypotheses, and action history, while Experience Memory $A=\{\mathrm{GP},S_1,\dots,S_n\}$9 retrieves clinical guidelines through AGRAG and historical cases via similarity-based importance estimation $P_{\mathrm{GP}}(x\mid q)<\tau$0 to propose critical missing evidence (Ge et al., 27 Mar 2026). Diagnosis is controlled by MCTS over actions $P_{\mathrm{GP}}(x\mid q)<\tau$1, with reward shaped by missing-evidence reduction and hypothesis-confidence increase (Ge et al., 27 Mar 2026). On MedChain, ClinicalAgents reports an average score of $P_{\mathrm{GP}}(x\mid q)<\tau$2, compared with $P_{\mathrm{GP}}(x\mid q)<\tau$3 for MedChain-Agents and $P_{\mathrm{GP}}(x\mid q)<\tau$4 for ReAct, along with test-ordering IoU $P_{\mathrm{GP}}(x\mid q)<\tau$5 versus $P_{\mathrm{GP}}(x\mid q)<\tau$6 for GPT-5.2, and image-examination DocLens $P_{\mathrm{GP}}(x\mid q)<\tau$7 versus $P_{\mathrm{GP}}(x\mid q)<\tau$8 (Ge et al., 27 Mar 2026).

Knowledge-graph grounding provides a more static but structured alternative. KG4Diagnosis combines a GP–specialist hierarchy with automated KG construction covering 362 common diseases, using semantic-driven entity and relation extraction, multi-dimensional decision relationship reconstruction, and human-guided knowledge expansion (Zuo et al., 2024). Specialist aggregation is represented as

$P_{\mathrm{GP}}(x\mid q)<\tau$9

while triage estimates $A=\mathrm{Clin}\cup \mathrm{Ex}$00 to route cases (Zuo et al., 2024). Although the paper emphasizes architectural guidelines more than benchmark superiority, it formalizes how explicit ontology-backed knowledge can constrain agent routing and downstream inference.

Self-learned knowledge is taken further in the MACD framework for reusable disease-specific experience. It decomposes learning into Knowledge Summarization, Knowledge Refinement, and Knowledge Application, with a Knowledge Summarizer extracting “General Criteria” and “Rare Criteria,” a Knowledge Refiner applying redundancy filtering via maximal marginal relevance and importance filtering via causal ablation, and a Diagnostician applying the resulting self-learned knowledge base $A=\mathrm{Clin}\cup \mathrm{Ex}$01 during inference (Li et al., 24 Sep 2025). On 4,390 MIMIC-IV and MIMIC-CDM cases across seven diseases, the self-learned system improves Llama-8B from $A=\mathrm{Clin}\cup \mathrm{Ex}$02 guideline-based accuracy to $A=\mathrm{Clin}\cup \mathrm{Ex}$03, DeepSeek-70B from $A=\mathrm{Clin}\cup \mathrm{Ex}$04 to $A=\mathrm{Clin}\cup \mathrm{Ex}$05, and Llama-70B from $A=\mathrm{Clin}\cup \mathrm{Ex}$06 to $A=\mathrm{Clin}\cup \mathrm{Ex}$07, with an overall average gain of $A=\mathrm{Clin}\cup \mathrm{Ex}$08 percentage points over clinical guidelines (Li et al., 24 Sep 2025). The same study reports that MACD-Human reaches $A=\mathrm{Clin}\cup \mathrm{Ex}$09 average accuracy versus $A=\mathrm{Clin}\cup \mathrm{Ex}$10 for Physicians-Only (Li et al., 24 Sep 2025).

MedAgentSim and EvoClinician offer memory in another form: case-derived exemplars and reflections. MedAgentSim stores correct consultation summaries in a Medical Records Buffer and reflective corrections in an Experience Records Buffer, retrieves similar cases via CLIP-based KNN, and ensembles multiple doctor agents by majority vote (Almansoori et al., 28 Mar 2025). EvoClinician stores action-context-outcome-grade tuples and evolves both rules and memory entries across cases (He et al., 30 Jan 2026).

A distinct but related contribution is causal structuring. MedCollab’s Hierarchical Disease Causal Chain transforms flattened diagnostic predictions into a DAG of pathological progression with explicit logical operators and optional conditional edge probabilities (Zhan et al., 1 Mar 2026). This suggests a shift from ranking diagnoses to modeling disease evolution and interdependence, potentially aligning MACD outputs with pathophysiological reasoning rather than surface-level label prediction.

6. Empirical performance, benchmarks, and failure modes

MACD research has produced a heterogeneous benchmark landscape, including rare-disease ranking, interactive case simulation, multimodal diagnostics, note-level problem detection, and free-text clinical prediction. Direct comparison across papers is limited by different datasets and task definitions, but several consistent empirical patterns emerge.

First, multi-agent systems often outperform corresponding single-agent baselines. On RareBench Combined, Mixed-Vendor MAC reports Recall@1/3/5/10 of $A=\mathrm{Clin}\cup \mathrm{Ex}$11, surpassing all single-LLM and Single-Vendor MAC configurations; on DiagnosisArena it reports Top-1 $A=\mathrm{Clin}\cup \mathrm{Ex}$12 and Top-5 $A=\mathrm{Clin}\cup \mathrm{Ex}$13, again the best among evaluated setups (Yuan et al., 14 Feb 2026). On MIMIC-Patient open-ended diagnosis, DynamiCare’s multi-agent GPT-4.1 outperforms single-agent GPT-4.1 across Hit@5, Hit@10, Rec@5, and Rec@10 (Shang et al., 3 Jul 2025). On MedChain, ClinicalAgents reports better overall average score than both single-agent and prior multi-agent baselines (Ge et al., 27 Mar 2026). MeDxAgent achieves a $A=\mathrm{Clin}\cup \mathrm{Ex}$14 accuracy gain over baseline and closes $A=\mathrm{Clin}\cup \mathrm{Ex}$15 of the gap to an oracle (Sanghvi et al., 2 Jun 2026). MedCollab reports higher diagnostic Accuracy by $A=\mathrm{Clin}\cup \mathrm{Ex}$16–$A=\mathrm{Clin}\cup \mathrm{Ex}$17 percentage points versus top agents, superior RaTEScore by $A=\mathrm{Clin}\cup \mathrm{Ex}$18–$A=\mathrm{Clin}\cup \mathrm{Ex}$19 percentage points in Diagnostic Basis and Differential Diagnosis, and measurable hallucination reduction via increased Entity-F1 (Zhan et al., 1 Mar 2026).

Second, diversity among agents matters. Mixed-vendor teams outperform single-vendor counterparts on both RareBench and DiagnosisArena, and overlap analysis shows that mixed-vendor teams “rescue” cases missed by the best single LLM: $A=\mathrm{Clin}\cup \mathrm{Ex}$20 on MME, $A=\mathrm{Clin}\cup \mathrm{Ex}$21 on HMS, $A=\mathrm{Clin}\cup \mathrm{Ex}$22 on LIRICAL, and $A=\mathrm{Clin}\cup \mathrm{Ex}$23 on DiagnosisArena (Yuan et al., 14 Feb 2026). The paper attributes this to complementary inductive biases and lower inter-vendor Jaccard similarity (Yuan et al., 14 Feb 2026). SOAP-note MAS work explicitly identifies “groupthink” as a failure mode and proposes heterogeneous LLMs as one mitigation direction (Lee et al., 29 Aug 2025).

Third, better components do not necessarily imply better systems. The Optimization Paradox study is explicit on this point: a Best-of-Breed system achieves $A=\mathrm{Clin}\cup \mathrm{Ex}$24 lab interpretation accuracy and stronger process metrics, yet only $A=\mathrm{Clin}\cup \mathrm{Ex}$25 diagnostic accuracy, compared with $A=\mathrm{Clin}\cup \mathrm{Ex}$26 for a top multi-agent system (Bedi et al., 6 Jun 2025). Error analysis shows hallucinated test results in $A=\mathrm{Clin}\cup \mathrm{Ex}$27 of BoB cases versus $A=\mathrm{Clin}\cup \mathrm{Ex}$28 in the top multi-agent system, unauthorized test ordering in $A=\mathrm{Clin}\cup \mathrm{Ex}$29 versus $A=\mathrm{Clin}\cup \mathrm{Ex}$30, and insufficient information gathering in $A=\mathrm{Clin}\cup \mathrm{Ex}$31 versus $A=\mathrm{Clin}\cup \mathrm{Ex}$32 (Bedi et al., 6 Jun 2025). This cautions against evaluating MACD solely by subtask metrics.

Fourth, interpretability gains are often reported alongside accuracy gains. Human evaluation in counterfactual multi-agent reasoning finds fewer factual errors ($A=\mathrm{Clin}\cup \mathrm{Ex}$33 vs. $A=\mathrm{Clin}\cup \mathrm{Ex}$34), fewer hallucinations ($A=\mathrm{Clin}\cup \mathrm{Ex}$35 vs. $A=\mathrm{Clin}\cup \mathrm{Ex}$36), lower critical harm risk ($A=\mathrm{Clin}\cup \mathrm{Ex}$37 vs. $A=\mathrm{Clin}\cup \mathrm{Ex}$38), and higher completeness ($A=\mathrm{Clin}\cup \mathrm{Ex}$39 vs. $A=\mathrm{Clin}\cup \mathrm{Ex}$40) compared with zero-shot CoT (You et al., 29 Mar 2026). MDTRoom’s clinician study reports improved ratings for tracking hypothesis evolution, understanding evidence usage, and identifying conflicts, with reduced mental demand and frustration (Kuai et al., 30 Mar 2026).

At the same time, important failure modes recur. Groupthink is directly documented in collaborative problem detection from SOAP notes, where a correct minority reader can flip to the majority’s incorrect view (Lee et al., 29 Aug 2025). Early anchoring appears in MeDxAgent, where differential questioning too early harms accuracy (Sanghvi et al., 2 Jun 2026). Hallucination remains a concern in simulated or weakly grounded environments such as MedAgentSim and interactive patient/examiner settings (Almansoori et al., 28 Mar 2025, He et al., 30 Jan 2026). Threshold sensitivity, token and latency cost, and prompt-format incompatibility also remain active issues (Ge et al., 27 Mar 2026, Bedi et al., 6 Jun 2025, Lu et al., 31 Mar 2026).

7. Human collaboration, privacy, and open research directions

Recent MACD work increasingly addresses deployment-relevant questions: how clinicians should inspect and intervene in agent reasoning, how cross-site evidence can be pooled under privacy constraints, and which design principles appear robust across tasks.

Human–AI collaboration is no longer treated only as a post hoc override. MDTRoom introduces an interactive workspace in which clinicians can add evidence, retract hypotheses, or guide individual agents, with interventions formalized as elements of

$A=\mathrm{Clin}\cup \mathrm{Ex}$41

and with updated states generated by $A=\mathrm{Clin}\cup \mathrm{Ex}$42 (Kuai et al., 30 Mar 2026). In a within-subject study with 12 clinicians, all participants preferred MDTRoom over the baseline transcript view (Kuai et al., 30 Mar 2026). MACD-human collaboration with self-learned knowledge escalates unresolved cases to a human physician after up to three rounds of evaluator-mediated agent consultation, reporting an $A=\mathrm{Clin}\cup \mathrm{Ex}$43 improvement over Physicians-Only diagnosis (Li et al., 24 Sep 2025). This suggests that hybrid workflows may be more realistic than fully autonomous MACD in near-term settings.

Privacy-preserving MACD has emerged as a distinct frontier. MedLatentDx addresses cross-hospital rare-disease diagnosis under privacy regulation by keeping patient records local and transmitting only compact latent KV blocks $A=\mathrm{Clin}\cup \mathrm{Ex}$44 to a host agent (Wang et al., 11 Jun 2026). In same-backbone settings it uses latent KV distillation; in heterogeneous settings it uses cross-family latent alignment (Wang et al., 11 Jun 2026). On CrossRare-Bench, MedLatentDx improves overall performance relative to Raw KV and LatentMAS baselines—for example, on a Llama host, accuracy $A=\mathrm{Clin}\cup \mathrm{Ex}$45, macro F1 $A=\mathrm{Clin}\cup \mathrm{Ex}$46, judge score $A=\mathrm{Clin}\cup \mathrm{Ex}$47—while reducing token-level reconstruction F1 to $A=\mathrm{Clin}\cup \mathrm{Ex}$48, an $A=\mathrm{Clin}\cup \mathrm{Ex}$49–$A=\mathrm{Clin}\cup \mathrm{Ex}$50 reduction relative to raw-KV baselines (Wang et al., 11 Jun 2026). This is notable because privacy concerns in MACD are not limited to explicit text sharing; even latent communication can leak content unless specifically controlled.

Several open directions are consistently identified across the literature. One is multimodality. LungNoduleAgent shows how MACD can be extended to CT-based malignancy reasoning through the sequence localizer $A=\mathrm{Clin}\cup \mathrm{Ex}$51 describer $A=\mathrm{Clin}\cup \mathrm{Ex}$52 reasoner, reporting Acc $A=\mathrm{Clin}\cup \mathrm{Ex}$53, F1 $A=\mathrm{Clin}\cup \mathrm{Ex}$54 on PrivateA, Acc $A=\mathrm{Clin}\cup \mathrm{Ex}$55, F1 $A=\mathrm{Clin}\cup \mathrm{Ex}$56 on PrivateB, and Acc $A=\mathrm{Clin}\cup \mathrm{Ex}$57, F1 $A=\mathrm{Clin}\cup \mathrm{Ex}$58 on LIDC-IDRI (Yang et al., 26 Nov 2025). MDAgents and consultation-flow RL frameworks also span text, image, and video settings (Kim et al., 2024, Wang et al., 19 Mar 2025). Another direction is formal orchestration, including learned routing, dynamic vendor selection, more granular uncertainty estimation, and stronger compatibility checks (Yuan et al., 14 Feb 2026, Ge et al., 27 Mar 2026, Bedi et al., 6 Jun 2025).

A recurring controversy concerns realism. Many evaluations use simulated patients, summarized records, or restricted benchmarks rather than prospective care environments. This suggests that current MACD results should be interpreted primarily as evidence about diagnostic reasoning architectures under controlled conditions, not as proof of clinical readiness. The literature itself repeatedly emphasizes clinician oversight, prospective validation, and calibration or deferral mechanisms before deployment (Kim et al., 2024, Almansoori et al., 28 Mar 2025, Sanghvi et al., 2 Jun 2026).

Taken together, current research positions MACD as a fast-evolving area at the intersection of LLM reasoning, clinical workflow modeling, and multi-agent systems. The field’s most technically distinctive contributions are dynamic recruitment, structured evidence grounding, interactive information collection, case-adaptive panel formation, memory-backed hypothesis verification, and transparent consensus protocols. The central unresolved question is no longer whether multiple agents can outperform a single model on some benchmarks, but which forms of specialization, communication, and oversight yield clinically reliable gains under realistic constraints (Zhan et al., 1 Mar 2026, Lu et al., 31 Mar 2026, Bedi et al., 6 Jun 2025).