A Review of Multi-Agent AI Systems for Biological and Clinical Data Analysis

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

LangGraph models a multi-agent system as a stateful graph, giving developers fine-grained control over execution paths.

“LangGraph, which models a multi-agent system as a stateful graph, giving developers fine-grained control over execution paths.”

• cites [14] Derouiche, H., Brahmi, Z., and Mezni, H. (2025). Agentic AI Frameworks: Architectures, Protocols, and Design Challenges. arXiv. sourcePartly supported

  The Comparative overview (p. 4) describes LangGraph as a graph-based model for sequencing tasks among LLM agents, with compositional flows and stateful operations. The Memory section (p. 5) also describes stateful graph nodes. The paper does not specifically attribute fine-grained developer control over execution paths to LangGraph; its explicit fine-grained-control description refers to Semantic Kernel. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

May combine two claims.

The virtual tumor board system leverages GPT-4 as a core reasoning engine and incorporates specialized precision oncology tools, including vision transformers for detecting microsatellite instability (MSI) and mutations (KRAS, BRAF) from histopathology slides, MedSAM for radiological image segmentation, and web-based search tools like OncoKB and PubMed.

“The system leverages GPT-4 as a core reasoning engine and incorporates specialized precision oncology tools, including vision transformers for detecting microsatellite instability (MSI) and mutations (KRAS, BRAF) from histopathology slides, MedSAM for radiological image segmentation, and web-based search tools like OncoKB and PubMed [34].”

• cites [34] Ferber, 2025, Development and validation of an autonomous artificial intelligence agent for clinical decision-making in oncology, Nat. Cancer sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Modern agent frameworks often implement a shared knowledge store or blackboard that agents can read from and write to during collaboration.

“Modern agent frameworks often implement a shared knowledge store or blackboard that agents can read from and write to during collaboration [16].”

• cites [16] Joshi, 2025, Review of autonomous systems and collaborative AI agent frameworks, Int. J. Sci. Res. Arch. sourcePartly supported

  Section 5.3, “Multi-Agent Systems for Collaborative Problem Solving” (page not provided), discusses collaborative frameworks and proposes that future systems enable agents to share knowledge. It does not say that modern frameworks often implement a shared store or blackboard that agents read from and write to. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim. May combine two claims.

MAS orchestration frameworks provide the infrastructure and design patterns to build systems where multiple AI agents collaborate and solve data-analysis problems more effectively than any individual agent alone.

“MAS orchestration frameworks provide the infrastructure and design patterns (messaging systems, shared memory stores, standardized agent interfaces, etc.) to build systems where multiple AI agents collaborate and solve data-analysis problems more effectively than any individual agent alone [9,16].”

• cites [9] Adimulam, A., Gupta, R., and Kumar, S. (2026). The Orchestration of Multi-Agent Systems: Architectures, Protocols, and Enterprise Adoption. arXiv. sourcePartly supported

  Sections III and IV describe orchestration infrastructure, communication protocols, and specialized agents working together; Section IV says their outcomes can surpass those of a single general-purpose agent. Section VIII gives a data-intensive insurance example, but the paper does not establish that orchestrated agents solve data-analysis problems generally more effectively than any individual agent alone. Page numbers are unavailable in our copy. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [16] Joshi, 2025, Review of autonomous systems and collaborative AI agent frameworks, Int. J. Sci. Res. Arch. sourcePartly supported

  Sections 2.1 and 5.3 (pages unavailable) describe orchestration frameworks and multi-agent systems for collaborative problem-solving. They do not establish that agents solve data-analysis problems more effectively than an individual agent, and the cited passages do not specify the messaging, shared-memory, or standardized-interface features in the printed sentence. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

In 2023, Google introduced an Agent-to-Agent (A2A) communication protocol as an open standard for LLM-based agents to exchange messages in a structured, secure manner.

“Notably, in 2023, Google introduced an Agent-to-Agent (A2A) communication protocol as an open standard for LLM-based agents to exchange messages in a structured, secure manner [25,26].”

• cites [25] Louck, Y., Stulman, A., and Dvir, A. (2025). Improving Google A2A Protocol: Protecting Sensitive Data and Mitigating Unintended Harms in Multi-Agent Systems. arXiv. sourcePartly supported

  The Introduction and Related Work (p. 1) describe Google’s A2A as a structured, secure framework for communication between agents. They do not call A2A an open standard, and the paper’s References (p. 12) date the cited announcement of the protocol to April 2025, not 2023. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [26] Surapaneni, R., Jha, M., Vakoc, M., and Segal, T. (2025, October 01). Announcing the Agent2Agent Protocol (A2A)—A New Era of Agent Interoperability. Google for Developers Blog. Available online: https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Various feedback and self-correction techniques have been introduced to improve LLM reliability, such as Reflexion and Self-Refine.

“various feedback and self-correction techniques have been introduced to improve LLM reliability, such as Reflexion and Self-Refine [21,22].”

• cites [21] Shinn, N., Labash, M.F., and Tran, T. (2023). Reflexion: Language agents with verbal reinforcement learning. arXiv. sourceSupported
• cites [22] Madaan, 2023, Self-refine: Iterative refinement with self-feedback, Adv. Neural Inf. Process. Syst. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Su et al. report relative improvements but do not specify exact percentage deltas; therefore, we treat this as illustrative, pending peer-reviewed, fully quantified results.

“It is important to note that Su et al. report relative improvements but do not specify exact percentage deltas; therefore, we treat this as illustrative, pending peer-reviewed, fully quantified results.”

• cites [32] Su, H., Long, W., and Zhang, Y. (2025). BioMaster: Multi-agent system for automated bioinformatics analysis workflows. bioRxiv. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Methods such as Toolformer showed that LLMs can learn to call external APIs to aid in problem-solving.

“methods such as Toolformer showed that LLMs can learn to call external APIs to aid in problem-solving [18].”

• cites [18] Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Hambro, E., and Scialom, T. (2023). Toolformer: Language models can teach themselves to use tools. arXiv. sourceSupported

May combine two claims.

In benchmarked evaluations, TxAgent achieved 92.1% accuracy on DrugPC (open-ended drug reasoning), and 93.6%/93.7% on BrandPC/GenericPC, outperforming GPT-4o and other tool-use LLMs across 3168 drug reasoning tasks and 456 personalized treatment scenarios.

“In benchmarked evaluations, TxAgent achieved 92.1% accuracy on DrugPC (open-ended drug reasoning), and 93.6%/93.7% on BrandPC/GenericPC, outperforming GPT-4o and other tool-use LLMs across 3168 drug reasoning tasks and 456 personalized treatment scenarios.”

• cites [31] Gao, S., Zhu, R., Kong, Z., Noori, A., Su, X., Ginder, C., and Zitnik, M. (2025). TxAgent: An AI agent for therapeutic reasoning across a universe of tools. arXiv. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

An analysis of the A2A framework identified potential attack vectors (e.g., prompt injection, token reuse, agent impersonation) and emphasized the need for agent authentication and permission controls.

“An analysis of the A2A framework identified potential attack vectors (e.g., prompt injection, token reuse, agent impersonation) and emphasized the need for agent authentication and permission controls [27].”

• cites [27] Zhang, Z., Cui, S., Lu, Y., Zhou, J., Yang, J., Wang, H., and Huang, M. (2024). Agent-SafetyBench: Evaluating the safety of multi-agent AI systems. arXiv. sourceCould not verify

  The Failure Mode Analysis (§4.3; page not provided) discusses tool-use robustness and risk awareness, not an analysis of the A2A framework. Searches for A2A, token reuse, agent impersonation, and authentication found no passage addressing the claimed attack vectors or proposed controls. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

An agent team can adopt a self-consistency approach where multiple agents independently work on the same problem and then vote or converge on an answer.

“an agent team can adopt a self-consistency approach where multiple agents independently work on the same problem and then vote or converge on an answer [19].”

• cites [19] Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., and Zhou, D. (2022). Self-consistency improves chain-of-thought reasoning in language models. arXiv. sourcePartly supported

  The Table 5 passage (section E.4; page not provided) reports a self-consistency variant using five responses per question, supporting the idea of multiple attempts at the same problem. It does not describe independent agents working as a team or say that they vote or converge on an answer. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

In an MAS context, guardrails can be implemented by assigning a dedicated critic or reviewer agent to evaluate and correct the outputs of other agents.

“In an MAS context, such guardrails can be implemented by assigning a dedicated critic or reviewer agent to evaluate and correct the outputs of other agents [23,24].”

• cites [23] Wang, T., Yu, P., Tan, X.E., O’Brien, S., Pasunuru, R., Dwivedi-Yu, J., and Celikyilmaz, A. (2023). Shepherd: A Critic for Language Model Generation. arXiv. sourcePartly supported

  The Introduction describes Shepherd as a dedicated model that critiques another model’s outputs, identifies errors, and suggests actionable refinements (page not specified). It does not describe an MAS guardrail implementation or show Shepherd directly correcting other agents’ outputs; those parts extend beyond the paper. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [24] He, 2025, Llm-based multi-agent systems for software engineering: Literature review, vision, and the road ahead, ACM Trans. Softw. Eng. Methodol. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Agents can implement a Tree-of-Thoughts strategy, collectively exploring different reasoning branches in parallel and exchanging information to decide on the best path forward.

“Agents can also implement a Tree-of-Thoughts strategy, collectively exploring different reasoning branches in parallel and exchanging information to decide on the best path forward [18,20].”

• cites [18] Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Hambro, E., and Scialom, T. (2023). Toolformer: Language models can teach themselves to use tools. arXiv. sourceCould not verify

  The Introduction describes a language model choosing when to use tools, and Section 4.2 describes its zero-shot evaluation. Neither those passages nor searches for Tree-of-Thoughts, parallel reasoning branches, and agents exchanging information found a passage addressing the attributed strategy. Our copy therefore does not verify this claim. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [20] Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T., Cao, Y., and Narasimhan, K. (2023). Tree of Thoughts: Deliberate Problem Solving with Large Language Models. arXiv. sourcePartly supported

  The Introduction and Section 5 describe Tree of Thoughts as exploring multiple reasoning paths and evaluating which to pursue. They describe language-model self-evaluation, not multiple agents collectively exploring branches or exchanging information; those parts of the attributed claim are unsupported. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

CrewAI organizes a multi-agent system as a team of role-based agents-Planner, Executor, Researcher, Reviewer, and Critic-who collaborate through structured but adaptive message exchanges.

“CrewAI, by contrast, organizes a multi-agent system as a team of role-based agents—Planner, Executor, Researcher, Reviewer, and Critic—who collaborate through structured but adaptive message exchanges.”

• cites [14] Derouiche, H., Brahmi, Z., and Mezni, H. (2025). Agentic AI Frameworks: Architectures, Protocols, and Design Challenges. arXiv. sourcePartly supported

  The Comparative overview (p. 3) describes CrewAI as promoting role-based collaboration, coordination, and delegation among agents. It does not identify the five named roles or describe CrewAI’s collaboration as structured but adaptive message exchanges. The Limitations section (p. 6) instead notes that predefined roles in CrewAI cannot easily change during execution. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

CrewAI leverages agent autonomy and role specialization to coordinate tasks dynamically based on the evolving problem state.

“Rather than enforcing a predefined graph, CrewAI leverages agent autonomy and role specialization to coordinate tasks dynamically based on the evolving problem state [14].”

• cites [14] Derouiche, H., Brahmi, Z., and Mezni, H. (2025). Agentic AI Frameworks: Architectures, Protocols, and Design Challenges. arXiv. sourcePartly supported

  The Comparative overview (p. 3) describes CrewAI as promoting role-based collaboration, coordination, and delegation, supporting the role-specialization part of the claim. But Limitations and Challenges (p. 6) says CrewAI’s predefined roles limit adaptability during execution and that agent interactions are often statically defined; it does not support the claim that CrewAI coordinates tasks dynamically as the problem evolves. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

Qualitatively, the dialogues revealed that agents effectively introduced diverse perspectives, such as identifying rare syndromes or flagging lab abnormalities that contradicted specific hypotheses, leading to a more thorough analysis than a single model's monologue.

“Qualitatively, the dialogues revealed that agents effectively introduced diverse perspectives, such as identifying rare syndromes or flagging lab abnormalities that contradicted specific hypotheses, leading to a more thorough analysis than a single model’s monologue.”

• cites [35] Chen, 2025, Enhancing diagnostic capability with multi-agent conversational LLMs, npj Digit. Med. sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

Token consumption for MAS reached over 6000 tokens compared to just 118 for standard input/output prompting.

“token consumption for MAS reached over 6000 tokens compared to just 118 for standard input/output prompting”

• cites [35] Chen, 2025, Enhancing diagnostic capability with multi-agent conversational LLMs, npj Digit. Med. sourcePartly supported

  The Results section, “Comparison with other methods to improve diagnostic performance” (page unavailable), says MAC produced the highest average output tokens, and the Methods section identifies Input/Output prompting as a comparison method. Our copy does not give the claimed figures of over 6,000 and 118 tokens, so the numerical comparison is unsupported. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

The quote only partly states this claim.

MAC's 34.11% primary accuracy beat Self-Refine (32.01%).

“Self-Refine (32.01%)”

• cites [35] Chen, 2025, Enhancing diagnostic capability with multi-agent conversational LLMs, npj Digit. Med. sourcePartly supported

  Table 1 in Results reports 34.11% accuracy for MAC’s most likely diagnosis in primary consultation. The Results comparison says MAC outperformed Self-Refine, but our copy does not give Self-Refine’s claimed 32.01%; it directs readers to Supplementary Table 12 for those results. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

May combine two claims.

The multi-agent approach significantly outperformed single-model baselines in both initial diagnosis and follow-up consultations.

“The multi-agent approach significantly outperformed single-model baselines in both initial diagnosis and follow-up consultations.”

• cites [35] Chen, 2025, Enhancing diagnostic capability with multi-agent conversational LLMs, npj Digit. Med. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

TrialGPT-Retrieval conducts a large-scale search to retrieve candidate trials from databases based on a patient's profile.

“TrialGPT-Retrieval, first conducts a large-scale search to retrieve candidate trials from databases based on a patient’s profile”

• cites [36] Jin, 2024, Matching patients to clinical trials with large language models, Nat. Commun. sourceSupported

May combine two claims.

In Chen et al.'s system, four AI doctor agents, each with a specialized knowledge focus or reasoning style, discuss a patient case, while a fifth agent acts as a supervisor to coordinate the conversation.

“four AI “doctor” agents, each with a specialized knowledge focus or reasoning style, discuss a patient case, while a fifth agent acts as a supervisor to coordinate the conversation”

• cites [35] Chen, 2025, Enhancing diagnostic capability with multi-agent conversational LLMs, npj Digit. Med. sourcePartly supported

  The Methods (Study design) and Results (Fig. 2 caption; pages unavailable) describe three doctor agents discussing a case with one coordinating supervisor, not four doctors and a fifth supervisor. The Methods (Subgroup analysis within MAC) describe assigning specialties as a tested variation; the Results say the standard configuration had no assigned specialties. The supervisor’s coordinating role is supported, but the claimed agent count and characterization of every doctor as specialized are not. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

May combine two claims.

Helpfulness rate for recommended further diagnostic tests reached 78.26%, an improvement over the GPT-4 baseline of 58.17% (p < 0.001).

“The helpfulness rate for recommended further diagnostic tests reached 78.26%, a marked improvement over the GPT-4 baseline of 58.17% (p < 0.001).”

• cites [35] Chen, 2025, Enhancing diagnostic capability with multi-agent conversational LLMs, npj Digit. Med. sourcePartly supported

  Table 1 in Results reports a 78.26% helpful rate for further diagnostic tests with the four-agent GPT-4 framework, versus 58.17% for standalone GPT-4, in primary consultations. The Results describe MAC’s improvement as significant, but our copy does not give p < 0.001 for this specific comparison; it also limits these figures to primary consultations. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

The quote only partly states this claim.

MAC's 34.11% primary accuracy beat Self-Consistency (32.45%).

“and Self-Consistency (32.45%) methods”

• cites [35] Chen, 2025, Enhancing diagnostic capability with multi-agent conversational LLMs, npj Digit. Med. sourcePartly supported

  In Results, Table 1 reports 34.11% primary-consultation accuracy for MAC with GPT-4 and four doctor agents. The same table assigns 32.45% to MAC with three doctor agents, not to Self-Consistency. The Results comparison says MAC outperformed Self-Consistency, but our copy does not substantiate the claimed 32.45% figure for Self-Consistency. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

May combine two claims.

The study's reliance on a small sample size (n = 20) and proprietary cloud-based models limits its generalizability and clinical robustness.

“the study’s reliance on a small sample size (n = 20) and proprietary cloud-based models suggests that future research must prioritize large-scale prospective evaluations and the development of local, open-source architectures”

• cites [34] Ferber, 2025, Development and validation of an autonomous artificial intelligence agent for clinical decision-making in oncology, Nat. Cancer sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The authors trained agents using 20,000 simulated patient encounters per task, evaluating them on a test set of 200 cases.

“The authors trained agents using 20,000 simulated patient encounters per task, evaluating them on a test set of 200 cases.”

• cites [39] Li, J., Lai, Y., Li, W., Ren, J., Zhang, M., Kang, X., and Liu, Y. (2025). Agent Hospital: A simulacrum of hospital with evolvable medical agents. arXiv. sourceSupported

May combine two claims.

This enables the agent to learn from experience and refine clinical reasoning over time.

“This enables the agent to learn from experience and refine clinical reasoning over time [39].”

• cites [39] Li, J., Lai, Y., Li, W., Ren, J., Zhang, M., Kang, X., and Liu, Y. (2025). Agent Hospital: A simulacrum of hospital with evolvable medical agents. arXiv. sourceSupported

May combine two claims.

Performance improvements arise specifically from the agent's ability to accumulate experience and leverage structured case memory, confirming that multi-agent interaction, not only the base model, drives accuracy gains in an experimental setting.

“This demonstrates that the performance improvements arise specifically from the agent’s ability to accumulate experience and leverage structured case memory—confirming that multi-agent interaction, not only the base model, drives the accuracy gains in an experimental setting.”

• cites [39] Li, J., Lai, Y., Li, W., Ren, J., Zhang, M., Kang, X., and Liu, Y. (2025). Agent Hospital: A simulacrum of hospital with evolvable medical agents. arXiv. sourceSupported

The quote only partly states this claim.

MedAgent-Zero significantly outperformed standard CoT (90.42%).

“and standard CoT (90.42%)”

• cites [39] Li, J., Lai, Y., Li, W., Ren, J., Zhang, M., Kang, X., and Liu, Y. (2025). Agent Hospital: A simulacrum of hospital with evolvable medical agents. arXiv. sourcePartly supported

  Table 4 reports 92.22% accuracy for MedAgent-Zero and 90.42% for CoT with GPT-4o on MedQA, supporting the comparison and the CoT figure. Section C.5 says MedAgent-Zero outperforms the other methods, but the cited passages do not establish that this 1.80-percentage-point difference is statistically significant. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

May combine two claims.

In the initial pass, TrialGPT-Retrieval successfully recalled over 90% of relevant trials while filtering out most irrelevant ones, requiring less than 6% of the total collection to maintain this high recall.

“TrialGPT-Retrieval successfully recalled over 90% of relevant trials while filtering out most irrelevant ones, requiring less than 6% of the total collection to maintain this high recall”

• cites [36] Jin, 2024, Matching patients to clinical trials with large language models, Nat. Commun. sourceSupported

May combine two claims.

MedAgent-Zero incorporates two memory systems: an experience base storing tens of thousands of simulated patient trajectories and a case base containing curated medical Q&A exemplars.

“MedAgent-Zero incorporates two memory systems: an experience base storing tens of thousands of simulated patient trajectories (symptoms → tests → diagnosis → outcome) and a case base containing curated medical Q&A exemplars.”

• cites [39] Li, J., Lai, Y., Li, W., Ren, J., Zhang, M., Kang, X., and Liu, Y. (2025). Agent Hospital: A simulacrum of hospital with evolvable medical agents. arXiv. sourcePartly supported

  The Doctor Agent Evolution section identifies both memory bases, and B.2 describes the medical case base as validated question–answer pairs. But Doctor Agent Evolution and B.3 say the experience base stores rules or principles learned from errors, not patient trajectories. C.2 reports treating 50,000 simulated patients; it does not say their trajectories were stored in that base. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

TrialGPT-Ranking assigns scores to rank the trials by suitability.

“TrialGPT-Ranking, assigns scores to rank the trials by suitability”

• cites [36] Jin, 2024, Matching patients to clinical trials with large language models, Nat. Commun. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

TrialGPT-Matching evaluates the patient's eligibility against each trial's criteria using LLM-based reading and annotation of inclusion/exclusion criteria.

“TrialGPT-Matching, then evaluates the patient’s eligibility against each trial’s criteria using LLM-based reading and annotation of inclusion/exclusion criteria”

• cites [36] Jin, 2024, Matching patients to clinical trials with large language models, Nat. Commun. sourceSupported

May combine two claims.

Current improvements may follow a 'scaling law' where gains slow as patient volume increases, suggesting an eventual performance ceiling.

“current improvements may follow a “scaling law” where gains slow as patient volume increases, suggesting an eventual performance ceiling”

• cites [39] Li, J., Lai, Y., Li, W., Ren, J., Zhang, M., Kang, X., and Liu, Y. (2025). Agent Hospital: A simulacrum of hospital with evolvable medical agents. arXiv. sourcePartly supported

  The “Scaling Laws in Evolution” section reports that diagnostic accuracy improved rapidly for the first 10,000 patient agents, then improved more slowly while continuing to rise through 50,000. This supports slowing gains as patient volume increases, but the paper does not suggest an eventual performance ceiling. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

In Agent-SafetyBench, MAS are challenged with scenarios and scored on rule adherence.

“In this benchmark, MAS are challenged with scenarios (making treatment recommendations under specific constraints) and scored on rule adherence.”

• cites [27] Zhang, Z., Cui, S., Lu, Y., Zhou, J., Yang, J., Wang, H., and Huang, M. (2024). Agent-SafetyBench: Evaluating the safety of multi-agent AI systems. arXiv. sourcePartly supported

  The Benchmark and Conclusion sections describe diverse test cases for LLM agents, and the Step4 passage says interactions receive safe-or-unsafe labels that determine a safety score. The paper does not describe the evaluated agents as multi-agent systems (MAS), and safety scoring is broader than rule adherence. It gives a medication-and-allergies example, but does not characterize the benchmark as treatment recommendations under specific constraints; no page numbers are available in our copy. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Zhang et al. (2024) introduced Agent-SafetyBench, a suite of tests to evaluate whether agents respect certain safety and ethical rules in medical and general contexts.

“Recently, Zhang et al., (2024) introduced Agent-SafetyBench, a suite of tests to evaluate whether agents respect certain safety and ethical rules in medical and general contexts [27].”

• cites [27] Zhang, Z., Cui, S., Lu, Y., Zhou, J., Yang, J., Wang, H., and Huang, M. (2024). Agent-SafetyBench: Evaluating the safety of multi-agent AI systems. arXiv. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

The socioeconomic burden of diagnostic errors is estimated to affect up to 1.8% of GDP in developed economies.

“While the architectural overhead of agentic orchestration is significant, the socioeconomic burden of diagnostic errors, which is estimated to affect up to 1.8% of GDP in developed economies, provides a strong argument for the “accuracy premium” offered by collaborative designs [55].”

• cites [55] Slawomirski, L., Kelly, D., de Bienassis, K., Kallas, K.A., and Klazinga, N. (2025). The Economics of Diagnostic Safety, OECD Publishing. OECD Health Working Papers, No. 176. sourcePartly supported

  Section 133 estimates the direct cost of diagnostic error at about 1.8% of GDP in a typical OECD country. The citing sentence calls this a socioeconomic burden, but the paper identifies the 1.8% figure as a direct healthcare cost; section 134 says indirect economic costs are difficult to estimate. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Academic groups have released research frameworks CAMEL and MetaGPT, providing templates for common MAS architectures.

“academic groups have released research frameworks, namely CAMEL and MetaGPT. These provide templates for common MAS architectures (e.g., collaborative coding agents, conversational QA teams) [13].”

• cites [13] Li, G., Hammoud, H., Itani, H., Khizbullin, D., and Ghanem, B. (2023). CAMEL: Communicative Agents for “Mind” Exploration of Large Language Model Society. arXiv. sourcePartly supported

  The Introduction describes CAMEL as an open-source cooperative-agent framework with modular agent implementations, and Section 3.1 shows agents collaborating through conversation on a coding task. These passages support CAMEL as an example of a research framework for a multi-agent architecture, but our copy does not establish the claim about MetaGPT. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

Expanded attack surfaces and emergent risky behaviors have been observed in autonomous agent experiments.

“Already, expanded attack surfaces and emergent risky behaviors have been observed in autonomous agent experiments [49,50,51].”

• cites [49] Deng, 2025, AI Agents Under Threat: A Survey of Key Security Challenges and Future Pathways, ACM Comput. Surv. sourceSupported
• cites [50] Sapkota, 2025, AI agents vs. agentic AI: A conceptual taxonomy, applications and challenges, Inf. Fusion sourcePartly supported

  Section 5.0.2 (page unavailable) says agentic AI has an expanded attack surface and that interactions among autonomous agents can produce unintended behavior and safety risks. It discusses these as challenges, but does not establish that both were observed in autonomous-agent experiments. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [51] Su, H., Luo, J., Liu, C., Yang, X., Zhang, Y., Dong, Y., and Zhu, J. (2025). A Survey on Autonomy-Induced Security Risks in Large Model-Based Agents. arXiv. sourceSupported

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Improving transparency and auditability of MAS reasoning will build confidence that these systems are working as intended.

“Improving transparency and auditability of MAS reasoning will build confidence that these systems are working as intended [69].”

• cites [69] Jimenez-Romero, C., Yegenoglu, A., and Blum, C. (2025). Multi-agent systems powered by large language models: Applications in swarm intelligence. Front. Artif. Intell., 8. sourcePartly supported

  Table 5 in the Conclusion identifies reduced transparency and limited explainability of agent decisions as weaknesses, and says unpredictable behavior requires careful monitoring. The paper does not say that improving auditability or transparency will build confidence that MAS are working as intended. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

May combine two claims.

Studies on physician attitudes toward AI have found that clear explanations and the ability to veto or override AI suggestions are important for acceptance.

“Studies on physician attitudes toward AI have found that clear explanations and the ability to veto or override AI suggestions are important for acceptance [59,60].”

• cites [59] Sullivan, 2019, Are Current Tort Liability Doctrines Adequate for Addressing Injury Caused by AI?, AMA J. Ethics sourceCould not verify

  The Abstract discusses the difficulty of explaining black-box AI recommendations in the context of liability, not physician acceptance. The section “Modify the standard of care” discusses clinicians evaluating AI results. Searches for physician attitudes, acceptance, and the ability to veto or override AI suggestions found no passage addressing the claimed findings; page numbers were not available in our copy. (LLM, conf=0.73)

  Experimental — accuracy not yet measured
• cites [60] Rigby, 2019, Ethical Dimensions of Using Artificial Intelligence in Health Care, AMA J. Ethics sourceCould not verify

  The editor’s article (page not provided) discusses physician input, AI as a complementary tool, and difficulty explaining black-box outputs. It does not report studies of physician attitudes or say that clear explanations and the ability to override AI suggestions affect acceptance. Searches for physician acceptance, explanations, vetoes, and overrides found no supporting passage. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

May combine two claims.

Recent evaluations indicate that structured agentic panels can reduce complex diagnostic expenditures from approximately $7850 to $2397 while significantly increasing specificity over human specialists.

“Recent evaluations indicate that structured agentic panels can reduce complex diagnostic expenditures from approximately $7850 to $2397 while significantly increasing specificity over human specialists [54,55].”

• cites [54] Klang, 2025, Assessing retrieval-augmented large language models for medical coding, NEJM AI sourceNot checked — no full text of the source
• cites [55] Slawomirski, L., Kelly, D., de Bienassis, K., Kallas, K.A., and Klazinga, N. (2025). The Economics of Diagnostic Safety, OECD Publishing. OECD Health Working Papers, No. 176. sourceCould not verify

  Sections 222 and 236 discuss possible diagnostic benefits of AI and potential cost savings, but neither reports an evaluation of structured agentic panels. Searches for “agentic panels,” $7,850, $2,397, and specificity relative to human specialists found no passage supporting the claimed results. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Ongoing work on federated learning and secure multi-party computation might inform how agents can operate on sensitive data without pooling it centrally.

“Ongoing work on federated learning and secure multi-party computation might inform how agents can operate on sensitive data without pooling it centrally [60].”

• cites [60] Rigby, 2019, Ethical Dimensions of Using Artificial Intelligence in Health Care, AMA J. Ethics sourceCould not verify

  The editor’s discussion of privacy and confidentiality in AI health care (Abstract; page not provided) does not address federated learning, secure multi-party computation, or operating on sensitive data without central pooling. Searches for those techniques and for privacy-preserving distributed learning found no passage in our copy that speaks to the claim. (LLM, conf=0.74)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The proposed Model Context Protocol (MCP) mediates agent access to external data with fine-grained security rules, acting as an intelligent broker that filters queries and enforces policies.

“for instance, the proposed Model Context Protocol (MCP) mediates agent access to external data with fine-grained security rules, acting as an intelligent broker that filters queries and enforces policies [52,53,64].”

• cites [52] Singh, A., Ehtesham, A., Kumar, S., and Khoei, T.T. (2025). A Survey of the Model Context Protocol (MCP): Standardizing Context to Enhance LLMs. Preprints. sourcePartly supported

  The Introduction describes MCP as connecting AI models to external resources through servers that securely access data, and the Security and Security & Validation sections describe authorization and controlled execution policies (pages not provided). These passages support mediated, policy-controlled access, but they do not describe MCP as an intelligent broker that filters queries or establish the claimed fine-grained security rules. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [53] Hou, X., Zhao, Y., Wang, S., and Wang, H. (2025). Model Context Protocol (MCP): Landscape, Security Threats, and Future Research Directions. arXiv. sourceNot checked — no full text of the source
• cites [64] IMDRF Software as a Medical Device (SaMD) Working Group (2025). Characterization Considerations for Medical Device Software and Software-Specific Risk. IMDRF/SaMD WG/N81 FINAL:2025, International Medical Device Regulators Forum.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

In clinical trial matching, multi-agent frameworks achieved 87.3% accuracy and enhanced clinician screening efficiency by 42.6% (p < 0.001).

“In clinical trial matching, multi-agent frameworks achieved 87.3% accuracy and enhanced clinician screening efficiency by 42.6% (p < 0.001).”

May combine two claims.

Many of the studies summarized in this review evaluate MAS on benchmark datasets or simulated scenarios rather than in live clinical environments, so their reported success may not directly translate to real-world settings where unpredictability is higher.

“Second, many of the studies we summarized evaluate MAS on benchmark datasets or simulated scenarios rather than in live clinical environments, which means their reported success may not directly translate to real-world settings where unpredictability is higher.”

May combine two claims.

Early evidence indicates that MAS can reduce oversights, catch inconsistencies, and provide clearer rationale through inter-agent dialogue.

“We see early evidence that they can reduce oversights, catch inconsistencies, and provide clearer rationale through inter-agent dialogue.”

Classified as probably not a substantial claim.

Studies were included only if they described the implementation of a multi-agent AI system where multiple autonomous agents collaborated or coordinated to perform tasks.

“To be included, a study needed to describe the implementation of a multi-agent AI system where multiple autonomous agents collaborated or coordinated to perform tasks.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The orchestration layer-not merely the agent design itself-is increasingly the critical determinant of system performance, safety, and clinical relevance.

“We argue that the orchestration layer—not merely the agent design itself—is increasingly the critical determinant of system performance, safety, and clinical relevance.”

Classified as probably not a substantial claim.

Studies focusing on single-LLM systems were excluded, except where such a system was integrated into a larger multi-agent framework.

“We excluded studies focusing on single-LLM systems, except where such a system was integrated into a larger multi-agent framework.”

Classified as probably not a substantial claim.

The search window was extended through 2026 via iterative manual screening of emerging preprints and recently published journal articles.

“we extended our search window through 2026 via iterative manual screening of emerging preprints and recently published journal articles.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

LangGraph maintains a persistent state throughout the workflow: intermediate results, memory, and context are stored directly in the graph's state object and passed deterministically across nodes.

“Unlike linear pipelines, LangGraph maintains a persistent state throughout the workflow: intermediate results, memory, and context are stored directly in the graph’s state object and passed deterministically across nodes.”

Probably not a substantial claim: may not pass S2 (checkable).

Clinical and research adoption of MAS requires the development of deterministic orchestration and rigorous cost-utility frameworks to ensure safety and expert-centered oversight.

“its clinical and research adoption requires the development of deterministic orchestration and rigorous cost-utility frameworks to ensure safety and expert-centered oversight”

Probably not a substantial claim: may not pass S3 (load-bearing).

These design choices make LangGraph particularly effective for structured, protocol-bound, or safety-critical applications where reproducibility and auditability matter.

“These design choices make LangGraph particularly effective for structured, protocol-bound, or safety-critical applications where reproducibility and auditability matter.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Compared to a single-agent setting, an MAS more realistically represents complex real-world scenarios involving multiple decision-makers or information sources.

“Compared to a single-agent setting, an MAS more realistically represents complex real-world scenarios involving multiple decision-makers or information sources [7,8].”

• cites [7] Wooldridge, 1995, Intelligent agents: Theory and practice, Knowl. Eng. Rev. sourceNot checked — no full text of the source
• cites [8] Stone, 2000, Multiagent systems: A survey from a machine learning perspective, Auton. Robot. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The initial search strategy focused on the preceding six-year window (approximately 2019-2025).

“The initial search strategy focused on the preceding six-year window (approximately 2019–2025)”

Probably not a substantial claim: may not pass S2 (checkable).

MAS enables a shift toward collaborative intelligence in biomedicine.

“MAS enables a shift toward collaborative intelligence in biomedicine”

Probably not a substantial claim: may not pass S3 (load-bearing).

The literature search spanned multiple databases: PubMed, arXiv, ACM, and IEEE Xplore.

“The search spanned multiple databases: PubMed, arXiv, ACM, and IEEE Xplore.”

Classified as probably not a substantial claim.

Studies that reported performance or utility improvements of the multi-agent approach over single-agent baselines were prioritized as evidence of the value added by agent collaboration.

“We prioritized studies that reported performance or utility improvements of the multi-agent approach over single-agent baselines as evidence of the value added by agent collaboration.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

A team of agents autonomously generating and executing complex action sequences raises the risk of propagating errors or undesirable behaviors.

“A team of agents autonomously generating and executing complex action sequences raises the risk of propagating errors or undesirable behaviors.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

MAS will require defense-in-depth: sandboxing of agent actions, well-defined scopes for each agent's authority, and continuous monitoring for anomalous behaviors.

“In general, MAS will require defense-in-depth: sandboxing of agent actions, well-defined scopes for each agent’s authority, and continuous monitoring for anomalous behaviors.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

Effective MAS requires mechanisms for memory sharing, safety constraints, and structured inter-agent communication.

“Effective MAS requires mechanisms for memory sharing, safety constraints, and structured inter-agent communication.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Drug discovery is a complex, multi-factorial process that can benefit from the division of labor offered by agentic workflows.

“Drug discovery is a complex, multi-factorial process that can benefit from the division of labor offered by agentic workflows.”

• AssumptionRests on an assumption: Drug discovery is a complex, multi-factorial process that can benefit from the division of labor offered by agentic workflows.

Probably not a substantial claim: may not pass S3 (load-bearing).

These differences reflect two distinct orchestration philosophies: graph-driven determinism versus role-driven collaboration.

“These differences reflect two distinct orchestration philosophies: graph-driven determinism versus role-driven collaboration.”

Probably not a substantial claim: may not pass S3 (load-bearing).

One safeguard is to include a specialized monitoring or moderation agent that watches the agents' interactions and intervenes if policies are violated.

“One safeguard is to include a specialized monitoring or moderation agent that watches the agents’ interactions and intervenes if policies are violated”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Alignment techniques from single-agent LLMs, such as content filtering and reinforcement learning from human feedback (RLHF), can be extended to MAS settings by applying them to each agent's outputs or decisions.

“Alignment techniques from single-agent LLMs, such as content filtering and reinforcement learning from human feedback (RLHF), can be extended to MAS settings by applying them to each agent’s outputs or decisions.”

Probably not a substantial claim: may not pass S3 (load-bearing).

MAS has rapidly gained traction as a strategy for tackling the complexity of biomedical data analysis and clinical decision-making.

“MAS has rapidly gained traction as a strategy for tackling the complexity of biomedical data analysis and clinical decision-making.”

• AssumptionRests on an assumption: MAS has rapidly gained traction as a strategy for tackling the complexity of biomedical data analysis and clinical decision-making.

Probably not a substantial claim: may not pass S4 (based).

As these frameworks evolve toward increasingly autonomous and distributed architectures, these same communication and control mechanisms introduce new vectors for error propagation, bias amplification, or unsafe tool invocation.

“As these frameworks evolve toward increasingly autonomous and distributed architectures, these same communication and control mechanisms introduce new vectors for error propagation, bias amplification, or unsafe tool invocation.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Techniques such as ReAct interleave logical Reasoning and concrete Acting steps by an LLM, demonstrating how language models can use external tools dynamically.

“Techniques, such as ReAct, interleave logical Reasoning and concrete Acting steps by an LLM, demonstrating how language models can use external tools dynamically [17].”

• cites [17] Yao, 2023, ReAct: Synergizing reasoning and acting in language models, Adv. Neural Inf. Process. Syst.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Ensuring secure and authenticated inter-agent communication will be vital so that a rogue agent cannot impersonate a trusted one or leak sensitive data.

“Ensuring secure and authenticated inter-agent communication will be vital so that a rogue agent cannot impersonate a trusted one or leak sensitive data.”

Probably not a substantial claim: may not pass S3 (load-bearing).

These paradigms are complementary rather than mutually exclusive.

“Importantly, these paradigms are complementary rather than mutually exclusive.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Shared memory facilitates experience replay, where agents can learn from prior cases by referencing how similar problems were solved, improving continual learning.

“Shared memory also facilitates experience replay, where agents can learn from prior cases by referencing how similar problems were solved, improving continual learning”

Probably not a substantial claim: may not pass S3 (load-bearing).

Shared memory ensures that facts discovered by one agent persist and inform the others.

“This ensures that facts discovered by one agent (e.g., a relevant patient finding or an experimental result) persist and inform the others.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

MAS in drug discovery is still in its infant stages, and most systems are prototypes or proofs-of-concept.

“MAS in drug discovery is still in its infant stages, and most systems are prototypes or proofs-of-concept.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The incorporation of a knowledge-graph agent helps reduce factual errors by grounding answers in curated databases, addressing a key limitation of standalone LLMs.

“Notably, the incorporation of a knowledge-graph agent helps reduce factual errors by grounding answers in curated databases, addressing a key limitation of standalone LLMs.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

In practice, LangGraph excels when workflows require strict branching logic, formal validation, or deterministic execution, while CrewAI shines when tasks benefit from adaptive teamwork and iterative critique.

“In practice, LangGraph excels when workflows require strict branching logic, formal validation, or deterministic execution, while CrewAI shines when tasks benefit from adaptive teamwork and iterative critique.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

This marks a shift toward a paradigm where AI acts as a digital coordinator of diverse medical resources.

“This marks a shift toward a paradigm where AI acts as a digital coordinator of diverse medical resources [28,29].”

• cites [28] Sutton, R.S., and Barto, A.G. (2018). Reinforcement Learning: An Introduction, MIT Press. [2nd ed.].Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [29] Ferber, D., Wei, J., Ghaffari Laleh, N., Wu, Z., Tan, Y., Peng, C., Wang, X., Liu, Y., Zhang, Z., and Chen, J. (2025). Multimodal Oncology Agent for IDH1 Mutation Prediction in Low-Grade Glioma. arXiv.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

MAS is naturally suited to multimodal reasoning because different agents can specialize in processing distinct data types and then collaborate to form a comprehensive assessment.

“MAS is naturally suited to such multimodal reasoning: different agents can specialize in processing distinct data types and then collaborate to form a comprehensive assessment.”

• AssumptionRests on an assumption: Clinical diagnosis often requires integrating information from multiple sources, specifically physical exam findings, laboratory results, and medical images.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

An AI research assistant was able to draft research proposals and survey prior work with minimal human input.

“an AI research assistant was able to draft research proposals and survey prior work with minimal human input [38]”

• cites [38] Naumov, V., Zagirova, D., Lin, S., Xie, Y., Gou, W., Urban, A., and Zhavoronkov, A. (2025). DORA AI Scientist: Multi-agent Virtual Research Team for Scientific Exploration Discovery and Automated Report Generation. bioRxiv. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims. The quote only partly states this claim.

Parallelizing agent operations and allowing asynchronous processing can help reduce latency.

“Parallelizing agent operations and allowing asynchronous processing can help.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Zheng et al. (2025) outlined strategies for incorporating lifelong learning into agent architecture by dividing the system into a perception module, a memory module, and an action module.

“Zheng et al. (2025) outlined strategies for incorporating lifelong learning into agent architecture by dividing the system into a perception module (to handle new modalities), a memory module (to accumulate evolving knowledge), and an action module (to adapt behaviors) [47].”

• cites [47] Zheng, J., Shi, C., Cai, X., Li, Q., Zhang, D., Li, C., Yu, D., and Ma, Q. (2025). Lifelong learning of large language model based agents: A roadmap. IEEE Trans. Pattern Anal. Mach. Intell., 1–20. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Latency is a concern for clinical applications.

“Latency is also a concern for clinical applications: doctors need answers fast.”

• AssumptionRests on an assumption: Doctors need answers fast.

Probably not a substantial claim: may not pass S4 (based).

Current frameworks are primarily limited by their focus on multiple-choice formats rather than the continuous, interactive narrative of patient-centered diagnostics.

“Current frameworks are primarily limited by their focus on multiple-choice formats rather than the continuous, interactive narrative of patient-centered diagnostics.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Microsoft has open-sourced Autogen as a Python 3.10+ library, which provides high-level abstractions for spawning agents, managing their message exchanges, and incorporating new agents dynamically.

“Microsoft has open-sourced Autogen (mentioned in Section 2.1) as a Python 3.10+ library, which provides high-level abstractions for spawning agents, managing their message exchanges, and incorporating new agents dynamically [12].”

• cites [12] Wu, Q., Bansal, G., Zhang, J., Wu, Y., Zhang, S., Zhu, E., Li, B., Jiang, L., Zhang, X., and Wang, C. (2023). AutoGen: Enabling next-gen LLM applications via multi-agent conversation frameworks. arXiv.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

If too many agents overlap in functionality, the system wastes time and computing resources on duplicate efforts.

“If too many agents overlap in functionality, the system wastes time and computing resources on duplicate efforts.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The system can iterate in a loop (plan → answer → review → refine) until a consensus or confidence threshold is reached.

“The system can iterate in a loop (plan → answer → review → refine) until a consensus or confidence threshold is reached.”

May combine two claims.

In complex medical coding tasks, a RAG-enhanced agentic system demonstrated superior performance to human providers; expert reviewers favored the AI's accuracy in 447 instances compared to 277 for provider-assigned codes (p < 0.001).

“For instance, in complex medical coding tasks where diagnostic specificity is tied to financial and legal outcomes, a RAG-enhanced agentic system demonstrated superior performance to human providers; expert reviewers favored the AI’s accuracy in 447 instances compared to 277 for provider-assigned codes (p < 0.001) [54].”

• cites [54] Klang, 2025, Assessing retrieval-augmented large language models for medical coding, NEJM AI sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

As the number of agents grows, communication overhead can grow combinatorially, a phenomenon sometimes called "agent flurry," where agents produce a torrent of messages that can slow down the system.

“As the number of agents grows, communication overhead can also grow combinatorially, a phenomenon sometimes called “agent flurry,” where agents produce a torrent of messages that can slow down the system.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

If two diagnostic agents disagree, that flags uncertainty for human review or triggers the system to gather more information.

“If two diagnostic agents disagree, that flags uncertainty for human review or triggers the system to gather more information.”

Probably not a substantial claim: may not pass S3 (load-bearing).

MCP acts as a safety layer-essentially a "secure USB port" through which AI agents can plug into critical data sources without breaching compliance or confidentiality.

“By enforcing standardized request formats, access controls (e.g., role-based permissions, user consent requirements), and comprehensive logging, MCP acts as a safety layer—essentially a “secure USB port” through which AI agents can plug into critical data sources without breaching compliance or confidentiality.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Most current MAS rely on foundation models that have a fixed knowledge cutoff and do not automatically update as new data arrives.

“Most current MAS rely on foundation models that have a fixed knowledge cutoff and do not automatically update as new data arrives.”

May combine two claims.

A retrieval agent might fetch irrelevant or misleading data, and a reasoning agent could incorporate it into a diagnosis, leading to an incorrect conclusion.

“a retrieval agent might fetch irrelevant or misleading data, and a reasoning agent could incorporate it into a diagnosis, leading to an incorrect conclusion”

Probably not a substantial claim: may not pass S3 (load-bearing).

If each agent needs the full history of the conversation or the entire patient record, the token usage can explode.

“One issue is the context length and memory: If each agent needs the full history of the conversation or the entire patient record, the token usage can explode.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

MCP is only a communication standard; thus, intra-system errors from agent reasoning are still a possibility.

“Finally, it is important to note MCP is only a communication standard; thus, intra-system errors from agent reasoning are still a possibility.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Agent role differentiation, specialized fine-tuning, or using smaller models for simpler agents can make the system leaner.

“Techniques such as agent role differentiation, perhaps via specialized fine-tuning for each agent’s subtask, or using smaller models for simpler agents can make the system leaner.”

Classified as probably not a substantial claim.

A clear distinction remains between simulation success and real-world bedside utility.

“a clear distinction remains between simulation success and real-world bedside utility”

Probably not a substantial claim: may not pass S4 (based).

Running several LLM agents in parallel or in long sequential dialogues can be resource-intensive in terms of CPU/GPU usage and memory.

“Running several LLM agents in parallel or in long sequential dialogues can be resource-intensive in terms of CPU/GPU usage and memory.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The shift from static, text-only models to evolvable and adaptive multi-agent ecosystems marks a pivotal transition in medical artificial intelligence.

“the shift from static, text-only models to evolvable and adaptive multi-agent ecosystems marks a pivotal transition in medical artificial intelligence”

Classified as probably not a substantial claim.

By moving beyond 'reading' medical literature to 'practicing' within synthetic simulations or dynamically assembling specialist teams, these systems begin to replicate the collaborative and iterative nature of real-world clinical decision-making.

“By moving beyond “reading” medical literature to “practicing” within synthetic simulations similar to Agent Hospital, or dynamically assembling specialist teams via MDAgents, these systems begin to replicate the collaborative and iterative nature of real-world clinical decision-making.”

Probably not a substantial claim: may not pass S3 (load-bearing).

After a Question-Answering Agent proposes an answer, a Reviewer Agent can assess its factual accuracy and coherence.

“after a Question-Answering Agent proposes an answer, a Reviewer Agent can assess its factual accuracy and coherence”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

It may be necessary to log detailed transcripts of agent dialogues and decision paths so that any failures can be audited after the fact.

“From a regulatory perspective, it may be necessary to log detailed transcripts of agent dialogues and decision paths so that any failures can be audited after the fact [57].”

• cites [57] (2025, July 05). High-Level Expert Group on Artificial Intelligence. Ethics Guidelines for Trustworthy AI. European Commission. Available online: https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Continual learning in deployed AI systems carries risks: models can forget old knowledge when fine-tuned on new data (catastrophic forgetting), or they can drift in undesirable ways.

“However, continual learning in deployed AI systems carries risks: models can forget old knowledge when fine-tuned on new data (catastrophic forgetting), or they can drift in undesirable ways (if new data is biased or if adversarial inputs are encountered).”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Lifelong learning for biomedical MAS is still largely uncharted territory.

“Lifelong learning for biomedical MAS is still largely uncharted territory.”

Classified as probably not a substantial claim.

With multiple agents contributing, it can be unclear which agent's action or which interaction led to an error.

“With multiple agents contributing, it can be unclear which agent’s action or which interaction led to an error.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Techniques similar to experience replay and periodic evaluation against fixed benchmarks can help detect when an agent's performance on prior knowledge starts degrading.

“Techniques similar to experience replay (re-training on a mix of old and new cases) and periodic evaluation against fixed benchmarks can help detect when an agent’s performance on prior knowledge starts degrading.”

May combine two claims.

A 2025 study developed a taxonomy of collaborative failure patterns in multi-agent medical systems, including flawed consensus driven by shared model biases, suppression of correct minority opinions, ineffective discussion dynamics, and critical information loss during synthesis.

“For example, a 2025 study developed a taxonomy of collaborative failure patterns, including flawed consensus driven by shared model biases, suppression of correct minority opinions, ineffective discussion dynamics, and critical information loss during synthesis [70].”

• cites [70] Gu, L., Zhu, Y., Sang, H., Wang, Z., Sui, D., Tang, W., Harrison, E., Gao, J., Yu, L., and Ma, L. (2025). MedAgentAudit: Diagnosing and Quantifying Collaborative Failure Modes in Medical Multi-Agent Systems. arXiv.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

Trust goes beyond raw accuracy: it involves human factors.

“This goes beyond raw accuracy: it involves human factors.”

May combine two claims.

The FDA has already signaled interest in adaptive AI that learns over time, proposing monitoring approaches to ensure safety is maintained during algorithm updates.

“The FDA has already signaled interest in adaptive AI that learns over time, proposing monitoring approaches to ensure safety is maintained during algorithm updates [58].”

• cites [58] Yang, G., Rao, A., Fernandez-Maloigne, C., Calhoun, V., and Menegaz, G. (2022, January 16–19). Explainable AI (XAI) In Biomedical Signal and Image Processing: Promises and Challenges. Proceedings of the 2022 IEEE International Conference on Image Processing (ICIP), Bordeaux, France. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable).

This enables human experts to focus more on empathy in patient care, insight in experimental design, or ethical judgment in difficult decisions.

“This enables human experts to focus more on empathy in patient care, insight in experimental design, or ethical judgment in difficult decisions.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The 'many hands' problem complicates assigning responsibility.

“This “many hands” problem complicates assigning responsibility.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

MASs often intensively share and process data, which raises privacy concerns.

“MASs often intensively share and process data, which raises privacy concerns.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The complexity that gives MAS power also makes them challenging to govern.

“The complexity that gives these systems power also makes them challenging to govern.”

May combine two claims.

Orchestrating multiple large language models is resource-intensive: context must be shared back-and-forth between agents, and they often engage in lengthy dialogues.

“Orchestrating multiple large language models is resource-intensive: context must be shared back-and-forth between agents, and they often engage in lengthy dialogues.”