TheraAgent: AI Systems for Therapeutic Tasks

• TheraAgent is a family of agentic AI systems designed for therapeutic decision-making and support through iterative, multi-step reasoning.
• These systems employ structured evaluation with specialists—such as planners, judges, and therapists—to ensure safety, accuracy, and evidence grounding in therapeutic outputs.
• Applications span mental health support, clinical treatment planning, drug discovery, and oncological theranostics, with performance validated by high user approval and rigorous metrics.

Searching arXiv for the cited TheraAgent-related papers to ground the article and citations. TheraAgent denotes a family of agentic AI systems for therapeutic tasks whose shared premise is that treatment support, therapeutic dialogue, or therapeutic design should not be handled as a single-pass text generation problem. Across recent work, the term has been used for at least four distinct but related paradigms: mental-health support chatbots such as TheraGen, explicitly framed as an AI therapy agent; orchestrated multi-agent systems for therapeutic discovery such as OrchestRA; inference-time treatment-planning systems built around iterative generate–judge–refine loops; and domain-specialized clinical reasoning agents for applications such as PET theranostics (Doshi et al., 2024, Suzuki et al., 25 Dec 2025, Li et al., 7 May 2026, Chen et al., 14 Mar 2026). Taken together, these systems characterize a “TheraAgent” as an AI architecture that couples therapeutic objectives with structured evaluation, tool use, memory, or evidence-grounded reasoning, typically under explicit safety constraints and with clear limits relative to licensed clinical care.

1. Conceptual scope and definitions

The most general use of the term refers to an AI agent for therapeutic decision-making or support that executes multi-step reasoning, rather than merely producing a one-shot response. In the therapeutic-design literature, OrchestRA defines a therapeutic multi-agent system as a coordinated set of specialist agents that plan, execute, and reason over target identification, molecular design or repositioning, and physiological or toxicity evaluation under a central Orchestrator (Suzuki et al., 25 Dec 2025). In clinical planning, TheraAgent is defined as an inference-time, self-improving therapeutic planning system that replaces one-shot LLM treatment generation with a deliberate workflow of generate–judge–refine guided by TheraJudge (Li et al., 7 May 2026). In mental-health support, a related framing appears in TheraGen, which is described as exemplifying an AI therapy agent by combining a transformer backbone with domain-specific fine-tuning, APA-grounded coping guidance, and deployment choices aimed at accessibility (Doshi et al., 2024). In PET theranostics, TheraAgent denotes a multi-agent, RAG-enabled framework for trial-calibrated pre-therapy prediction in metastatic castration-resistant prostate cancer (Chen et al., 14 Mar 2026).

These variants differ in domain and implementation, but converge on several design commitments. First, therapeutic outputs are treated as high-stakes artifacts requiring explicit verification, revision, or evidence calibration. Second, multi-dimensional quality—such as accuracy, completeness, safety, empathy, or pharmacokinetic viability—is operationalized as an internal control signal rather than a passive evaluation metric. Third, specialist decomposition is common: systems divide tasks among planners, judges, biologists, chemists, pharmacologists, critics, coaches, therapists, radiologists, biochemists, or oncologists (Suzuki et al., 25 Dec 2025, Rahman et al., 29 Jun 2026, Chen et al., 14 Mar 2026). This suggests that “TheraAgent” is best understood not as a single model architecture, but as an agentic design pattern for therapeutic reasoning.

A common misconception is that all TheraAgent systems are chatbots. The literature does not support that simplification. Some systems are conversational and user-facing, such as TheraGen (Doshi et al., 2024); others are optimization engines for drug discovery (Suzuki et al., 25 Dec 2025), treatment-planning frameworks for clinician-style regimen design (Li et al., 7 May 2026), or oncology decision-support systems centered on imaging, laboratory, and trial evidence (Chen et al., 14 Mar 2026).

2. Agentic architectures and control loops

A central architectural distinction in this literature is the move from direct generation to controlled iterative refinement. In treatment planning, the core loop is explicit: a Planner generates a candidate plan, TheraJudge evaluates it along clinically relevant dimensions, Memory stores plan–feedback–score triplets, and the loop continues until early stopping or a maximum iteration count, with the final output selected from the last few iterations by score (Li et al., 7 May 2026). The formalization uses a case $P = (d, s, y)$ and plan quality

$Q(T \mid P) = \sum_{i=1}^{N} q_i(T \mid P),$

where dimensions include Accuracy, Targeting, Completeness, and Harm Control or Safety (Li et al., 7 May 2026). The Planner and Judge are written as

$(T_k, c_k) = f_\theta(P, M^{(k-1)}),$

and

$(R_k, \{q_{k,i}\}_{i=1}^N, s_k) = g_\phi(P, T_k),$

with selection by aggregate score (Li et al., 7 May 2026).

The mental-health multi-agent formulation in a later TheraAgent paper is similarly iterative, but role-separated. TheraJudge evaluates a response across seven psychological dimensions; a Critic diagnoses low-scoring dimensions; a Coach converts those diagnoses into dimension-specific revision instructions; and a Therapist rewrites the response (Rahman et al., 29 Jun 2026). The loop iterates up to 8 times with early stopping when scores stabilize or exceed adequacy thresholds, and the best-scoring response is returned (Rahman et al., 29 Jun 2026). The evaluator is formalized as

$E_\theta(c, r) = (y_1,\dots,y_7),$

and the refinement operator as

$G_\phi(c, r, E_\theta(c, r)) \to r'.$

This architecture makes evaluation actionable by inserting it directly into the inference loop (Rahman et al., 29 Jun 2026).

OrchestRA uses a different control mechanism: a central Orchestrator implemented with LangGraph, described as a finite-state controller and stateful planner. It translates natural-language intent into structured workflows, maintains a global AgentState memory, enforces deterministic routing, validates artifacts such as SMILES syntax, and runs an iterative optimization loop driven by explicit approval flags rather than probabilistic LLM routing (Suzuki et al., 25 Dec 2025). Its sequential progression is Biologist $\rightarrow$ Chemist $\rightarrow$ Pharmacologist, with iterative returns from Pharmacologist feedback to Chemist optimization until approval or a maximum-iteration cap (Suzuki et al., 25 Dec 2025).

TxAgent also adopts a planner–executor structure, though centered on tool use. Each reasoning step $R_i$ contains a thought $T_i$, a set of function calls $Q(T \mid P) = \sum_{i=1}^{N} q_i(T \mid P),$0, and tool responses $Q(T \mid P) = \sum_{i=1}^{N} q_i(T \mid P),$1, and inference continues until the model emits a special [FinalAnswer] token and calls the Finish tool (Gao et al., 14 Mar 2025). Its formalism includes

$Q(T \mid P) = \sum_{i=1}^{N} q_i(T \mid P),$2

for contextual thought generation, and

$Q(T \mid P) = \sum_{i=1}^{N} q_i(T \mid P),$3

for function-call generation (Gao et al., 14 Mar 2025). This gives TheraAgent-like behavior through explicit reasoning traces, tool selection, and structured finalization.

3. Major application domains

The term spans at least four substantive therapeutic domains.

First, mental-health support systems. TheraGen is an AI-powered mental health chatbot based on Meta’s LLaMA 2 7B chat model, fine-tuned on approximately 1,000,000 conversational entries drawn from anonymized therapy transcripts, Reddit, Twitter, and APA literature (Doshi et al., 2024). It is explicitly positioned as a complementary tool rather than a replacement for professional therapy, provides empathetic responses and APA-informed coping strategies, and is deployed through a Flask-based web interface with Replicate-backed cloud inference (Doshi et al., 2024). A separate 2026 framework advances this direction by introducing TheraJudge and a Critic–Coach–Therapist loop for human-aligned response refinement in mental health support (Rahman et al., 29 Jun 2026).

Second, clinical treatment planning. The 2026 TheraAgent system for precise and comprehensive treatment planning is built around iterative generate–judge–refine, with DeepSeek-R1 used for both Planner and Judge in the main configuration (Li et al., 7 May 2026). Its target artifact is not a conversational reply but a therapeutic regimen that must specify medication choice, dose, route, indication, supportive care, monitoring, follow-up, and safety constraints (Li et al., 7 May 2026). This use of “TheraAgent” is narrower and more clinical than the mental-health usage.

Third, therapeutic discovery and molecular design. OrchestRA presents a human-in-the-loop autonomous multi-agent platform that unifies biology, chemistry, and pharmacology into an end-to-end therapeutic-design workflow (Suzuki et al., 25 Dec 2025). The Biologist reasons over a manually curated biomedical knowledge graph with 147,814 nodes and 13,957,458 edges; the Chemist performs pocket detection, de novo generation, docking, scoring, novelty analysis, and BO+GA optimization; the Pharmacologist runs ADMET prediction and a 5-compartment PBPK simulation (Suzuki et al., 25 Dec 2025). Here, “therapeutic” refers to candidate molecule design rather than patient-facing care.

Fourth, evidence-grounded oncology prediction. In PET theranostics, TheraAgent is a multi-agent framework for predicting response to $Q(T \mid P) = \sum_{i=1}^{N} q_i(T \mid P),$4-PSMA radioligand therapy in mCRPC using PSMA PET/CT reports, laboratory data, and clinical records (Chen et al., 14 Mar 2026). It combines multi-expert extraction, confidence-weighted consensus, self-evolving memory, and RAG over a 23-publication knowledge base centered on VISION and TheraP (Chen et al., 14 Mar 2026). This suggests that TheraAgent can also denote a domain-specific clinical decision-support architecture rather than a general therapeutic planner.

4. Core computational motifs

Despite domain diversity, several computational motifs recur.

Iterative self-improvement is the most visible. Treatment-planning TheraAgent uses Memory to store prior drafts and critiques, retrieving the Top–N = 3 items by score for the next iteration (Li et al., 7 May 2026). The mental-health refinement system likewise keeps the best-scoring response seen across up to 8 iterations and explicitly seeks to improve weak dimensions without degrading strong ones (Rahman et al., 29 Jun 2026).

Evaluation as a control variable is another defining motif. In treatment planning, TheraJudge outputs dimension-wise scores $Q(T \mid P) = \sum_{i=1}^{N} q_i(T \mid P),$5 and an aggregate score

$Q(T \mid P) = \sum_{i=1}^{N} q_i(T \mid P),$6

with optional safety penalties for unsafe rule violations (Li et al., 7 May 2026). In the mental-health evaluator, the reward used during judge training is

$Q(T \mid P) = \sum_{i=1}^{N} q_i(T \mid P),$7

with weights $Q(T \mid P) = \sum_{i=1}^{N} q_i(T \mid P),$8 ordered as Guidance, Informativeness, Relevance, Safety, Empathy, Helpfulness, Understanding (Rahman et al., 29 Jun 2026). The conceptual shift is from “judge after generation” to “judge inside generation.”

Memory and retrieval appear in multiple forms. Treatment-planning TheraAgent stores $Q(T \mid P) = \sum_{i=1}^{N} q_i(T \mid P),$9 triples (Li et al., 7 May 2026). The PET theranostics system stores memory entries $(T_k, c_k) = f_\theta(P, M^{(k-1)}),$0 and retrieves top-$(T_k, c_k) = f_\theta(P, M^{(k-1)}),$1 similar cases based on prognostic features such as PSMA expression level, liver or lung metastasis, and prior chemotherapy (Chen et al., 14 Mar 2026). TxAgent uses a different retrieval target: ToolRAG retrieves the top-$(T_k, c_k) = f_\theta(P, M^{(k-1)}),$2 relevant tools by embedding similarity against tool descriptions (Gao et al., 14 Mar 2025).

Knowledge grounding and evidence calibration are especially prominent in systems that operate under rapidly changing biomedical evidence. TxAgent retrieves live information from openFDA and Open Targets, allowing it to answer with current approvals and contraindications rather than model-cutoff knowledge (Gao et al., 14 Mar 2025). OrchestRA grounds target discovery in a manually curated knowledge graph and later stages in docking, deep-learning scoring, ADMET models, and PBPK simulation (Suzuki et al., 25 Dec 2025). PET theranostic TheraAgent retrieves passages from VISION, TheraP, systematic reviews, and related outcome studies, then asks an LLM reasoner to synthesize case evidence and trial evidence with explicit citations (Chen et al., 14 Mar 2026).

Specialist decomposition with deterministic coordination recurs across systems. A plausible implication is that therapeutic tasks are being treated as composite workflows in which interpretability and error localization benefit from role separation. When a system distinguishes evaluator, critic, coach, and therapist (Rahman et al., 29 Jun 2026), or biologist, chemist, and pharmacologist (Suzuki et al., 25 Dec 2025), failures can be attributed to a more specific subsystem than in monolithic generation.

5. Data, knowledge sources, and tooling ecosystems

The underlying data and tool environments vary sharply by application.

TheraGen is trained on approximately 1,000,000 conversational entries, composed of 700,000 anonymized therapy session transcripts from Kaggle and Hugging Face, 100,000 Reddit posts and comments from mental-health-related subreddits, 10,000 tweets containing mental health hashtags, and excerpts from APA literature and clinical guidelines extracted via PyMuPDF (Doshi et al., 2024). Preprocessing includes lowercasing, removal of special characters and emojis, LLaMA 2 tokenization to a maximum sequence length of 512 tokens, regex-based anonymization, removal of incomplete or offensive entries, and filtering to clinically relevant APA sections (Doshi et al., 2024).

TxAgent’s knowledge environment is centered on ToolUniverse, which consolidates 211 biomedical tools linked to trusted sources such as openFDA, Open Targets, Human Phenotype Ontology, and PrimeKG (Gao et al., 14 Mar 2025). openFDA provides US FDA drug labeling data since 1939 and more than 67,000 marketed drugs; Open Targets contributes 63,121 targets, 28,327 diseases, and 18,041 drugs, with over 8M target–disease associations (Gao et al., 14 Mar 2025). The system is trained on 378,027 instruction-tuning samples derived from 85,340 multi-step reasoning traces containing 177,626 steps and 281,695 function calls (Gao et al., 14 Mar 2025).

OrchestRA’s biomedical knowledge graph integrates 13 primary data sources—Bgee, CTD, DrugBank, DrugCentral, GO, HPO, MONDO, NCBI Entrez Gene, Reactome, SIDER, UBERON, UMLS, and a human PPI network—into a KuzuDB property graph with standardized identifiers and 13,957,458 curated edges (Suzuki et al., 25 Dec 2025). In chemistry, it uses PDBFixer, OpenBabel, AutoDock Tools, P2Rank v2.4.1, AutoDock Vina v1.2, RDKit, DiffSBDD, Boltz-2, PyMOL, and MOE (2024.0601); in pharmacology, ADMET-AI, PySB, and SciPy are used (Suzuki et al., 25 Dec 2025).

The treatment-planning TheraAgent uses a curated corpus of more than 600 guideline documents for RAG inside TheraJudge, though RAG is disabled on HealthBench to avoid region-specific bias and enabled on real cases and ablations (Li et al., 7 May 2026). Its memorizer stores plan–feedback–score triples, retrieves the best three prior items by score, and uses those to condition the next draft (Li et al., 7 May 2026).

The PET theranostic system uses no supervised parameter training; instead, GPT-4o acts as the base LLM across agents, and the system relies on prompt engineering, SEA-Mem, and RAG over a curated knowledge base of 23 Lu-177 PSMA publications (Chen et al., 14 Mar 2026). To supplement 35 real mCRPC cases, it uses 400 synthetic cases generated from VISION and TheraP statistics, with complexity strata of clear (50%), ambiguous (30%), and misleading (20%) (Chen et al., 14 Mar 2026).

6. Quantitative performance and empirical findings

Reported performance varies by domain and metric, but the literature consistently presents TheraAgent systems as outperforming one-shot or less structured baselines.

For mental-health chatbot deployment, TheraGen reports that 94% of users reported improved mental well-being, 93% had a positive experience, and 90% would recommend the system (Doshi et al., 2024). Automatic metrics include BLEU = 0.67, ROUGE = 0.62, coherence score = 0.78, Distinct-1 = 0.82, and Distinct-2 = 0.76, with an average response time of 1395 ms (Doshi et al., 2024). Expert or human evaluation judged 87% of responses appropriate or accurate, and experts rated 92% as contextually appropriate or coherent (Doshi et al., 2024).

For therapeutic response refinement in mental health, TheraJudge achieves ICC(C,1) = 0.879–0.989 and ICC(A,1) = 0.848–0.981 across dimensions, with the paper stating ICC = 0.87–0.95 in the abstract (Rahman et al., 29 Jun 2026). Acting on those evaluations, the full TheraAgent yields a +0.43 improvement in human-rated therapeutic quality on a 5-point scale, from 4.26 to 4.69, with $(T_k, c_k) = f_\theta(P, M^{(k-1)}),$3 (Rahman et al., 29 Jun 2026). Low-quality responses ($(T_k, c_k) = f_\theta(P, M^{(k-1)}),$4) improve by +2.45 points on average, with a 94% recovery rate to $(T_k, c_k) = f_\theta(P, M^{(k-1)}),$5, and unsafe cases show 100% recovery for the Safety dimension among initially unsafe responses (Rahman et al., 29 Jun 2026). Dimension-level mean improvements are Helpfulness +0.73, Informativeness +0.63, Empathy +0.57, Guidance +0.40, Understanding +0.34, Relevance +0.28, and Safety +0.03, the latter reflecting ceiling effects (Rahman et al., 29 Jun 2026).

For treatment planning, the 2026 TheraAgent reports state-of-the-art automatic evaluation on HealthBench, with an overall score of 48.94, surpassing the second-best by 4.66 points (Li et al., 7 May 2026). It leads in Accuracy and Completeness by 2.91 and 4.43 points over the second-best system, and shows improvements of +6.78 in Hedging and 44.65 in Context Seeking (Li et al., 7 May 2026). In blinded real-world expert evaluations, it attains an 86% win rate against physicians, with strong wins in Targeting (69%), Completeness (71%), and Harm Control (51%); in three-way preference, it is “Most Preferred” in 65.7% of cases, versus 25.7% for the base model and 8.6% for physicians (Li et al., 7 May 2026). Agreement between TheraJudge and the official HealthBench evaluator is reported as Pearson correlation 0.7052, Spearman 0.6669, and CCC 0.6467 (Li et al., 7 May 2026).

For tool-grounded therapeutic reasoning, TxAgent achieves 93.8% accuracy on DrugPC multiple-choice and 92.1% on DrugPC open-ended tasks, compared with GPT-4o at 76.4% and 66.3% respectively (Gao et al., 14 Mar 2025). On TreatmentPC, it scores 86.8% multiple-choice and 75.0% open-ended; this open-ended score exceeds GPT-4o’s multiple-choice score of 74.1% (Gao et al., 14 Mar 2025). The system is also reported to outperform DeepSeek-R1 (671B) by 10.3% on TreatmentPC multiple-choice and 7.5% on TreatmentPC open-ended (Gao et al., 14 Mar 2025). Representation robustness across DrugPC, BrandPC, and GenericPC is summarized by a variance of 0.00667 for TxAgent, compared with 9.96 for GPT-4o (Gao et al., 14 Mar 2025).

For therapeutic design, OrchestRA’s ABL1 benchmark reports that DiffSBDD yielded 682 chemically valid candidates and that enrichment for repositioning reached EF@1% = 50.0%, with Axitinib and Dasatinib among the top-ranked compounds (Suzuki et al., 25 Dec 2025). In its end-to-end diabetes campaign, the final lead is reported as COC1CC(O)(c2ccncc2)CON1CC(=O)O, with PBPK predicting favorable plasma profiles after a single oral 200 mg dose (Suzuki et al., 25 Dec 2025). The paper also states that the pharmacologist approval concluded the optimization loop and that no significant increase in carcinogenicity or DILI risk was observed versus a random baseline (Suzuki et al., 25 Dec 2025).

For PET theranostics, TheraAgent achieves 75.7% overall accuracy and F1 82.1% on 35 real patients, with PSA-response accuracy 77.1% and OS$(T_k, c_k) = f_\theta(P, M^{(k-1)}),$6m accuracy 74.3% (Chen et al., 14 Mar 2026). On 400 synthetic cases, it reports 95.0% accuracy on clear cases, 79.2% on ambiguous cases, 78.8% on misleading cases, and 87.0% overall (Chen et al., 14 Mar 2026). Compared with MDAgents and MedAgent-Pro, the paper reports improvements exceeding 20% in the abstract, while the detailed results state +12.8 points over MDAgents and +18.6 points over MedAgent-Pro in overall accuracy on real data (Chen et al., 14 Mar 2026). This suggests the abstract-level comparison may refer to a broader or rounded comparison framework, whereas the detailed table values are more precise for the real clinical set.

7. Safety, limitations, and future directions

Safety occupies different roles across the TheraAgent literature, but it is rarely absent.

In mental-health support, TheraGen emphasizes transparency that it is an AI assistant, not a substitute for licensed care; anonymity without login; no retention of personal data; encrypted communications with Replicate; content filtering; trigger warnings; and bias mitigation through diverse training and regular audits (Doshi et al., 2024). However, crisis-handling or escalation protocols beyond content filtering and trigger warnings are not detailed (Doshi et al., 2024). The later mental-health TheraAgent strengthens this dimension by embedding Safety directly into both evaluator weighting and refinement policy. When Safety is below threshold, the Coach instructs the Therapist to avoid diagnostic or medical claims, remove harmful coping suggestions, add grounding and de-escalation steps, and suggest professional or emergency support when warranted (Rahman et al., 29 Jun 2026).

In treatment planning, safety is encoded in TheraJudge’s rubric and may also be modeled through penalties for unsafe plans:

$(T_k, c_k) = f_\theta(P, M^{(k-1)}),$7

where $(T_k, c_k) = f_\theta(P, M^{(k-1)}),$8 denotes rule-based safety checks such as contraindicated drugs in pregnancy, dosing above maximum, or severe interactions (Li et al., 7 May 2026). The judging rubric operationalizes checks such as DoseWithinRange, CheckContraindications, CheckDDI, RequireMonitoring, IndicationCriteriaMet, and AntibioticStewardship (Li et al., 7 May 2026). This use of judgment as an internal safety mechanism distinguishes the framework from systems that merely filter outputs after generation.

In therapeutic design, OrchestRA’s Pharmacologist evaluates DILI, carcinogenicity, and hERG risk, and feeds these diagnostics back into molecular optimization (Suzuki et al., 25 Dec 2025). Still, the paper explicitly notes that explicit off-target binding prediction is not included in this version, and that the system remains purely in silico and dependent on the fidelity of docking, Boltz-2, and ADMET-AI predictions (Suzuki et al., 25 Dec 2025).

In PET theranostics, safety is framed mainly as hallucination mitigation and evidence grounding. Predictions are constrained by multi-expert consensus, self-evolving memory, and trial citations from VISION and TheraP (Chen et al., 14 Mar 2026). Yet the authors acknowledge several limitations: the real dataset is small ($(T_k, c_k) = f_\theta(P, M^{(k-1)}),$9), synthetic data may not capture all real-world heterogeneity, no explicit probabilistic calibration is implemented, and the Radiologist Agent parses reports rather than raw PET/CT volumes (Chen et al., 14 Mar 2026).

Across the literature, several broader limitations recur. Inference-time cost is substantial in iterative systems. Treatment-planning TheraAgent reports that 3 iterations require about 6 calls, 13,445 tokens, and about 333 s, or 9.9× the cost of a single DeepSeek-R1 generation, while 10 iterations require about 20 calls, 87,005 tokens, and about 754 s, or 64.1× cost (Li et al., 7 May 2026). Backbone dependence remains unresolved: the planning system is validated on strong LLMs such as DeepSeek-R1, GPT-4o, and o4-mini, but behavior on smaller LLMs is not fully characterized (Li et al., 7 May 2026). Guideline heterogeneity and jurisdictional variation complicate RAG use, as shown by the decision to disable RAG on HealthBench because it degraded performance there (Li et al., 7 May 2026). Data scarcity and domain shift remain central in specialized areas such as theranostics (Chen et al., 14 Mar 2026) and mental health (Rahman et al., 29 Jun 2026).

Future directions are correspondingly diverse. TheraGen proposes telehealth integration, speech recognition, multilingual support, and long-term monitoring (Doshi et al., 2024). Treatment-planning TheraAgent points toward multimodal integration of labs time series, imaging, and monitors, as well as multi-morbidity and longitudinal care extensions (Li et al., 7 May 2026). OrchestRA proposes physical integration with self-driving labs, broader ADMET endpoints, richer toxicity modeling, continual knowledge-graph updates, multi-omics integration, and multi-agent-to-robotics coupling (Suzuki et al., 25 Dec 2025). PET theranostic TheraAgent calls for multi-center datasets, reinforcement learning from clinical feedback, and prospective validation (Chen et al., 14 Mar 2026).

A plausible implication is that future uses of the term “TheraAgent” will continue to bifurcate into at least two lines: therapeutic communication systems centered on alignment, empathy, and safety; and therapeutic engineering or clinical-decision systems centered on evidence-grounded planning, tool use, and structured optimization. The existing literature supports both trajectories, but also makes clear that none of these systems is presented as a replacement for licensed clinicians, wet-lab validation, or institutional clinical governance (Doshi et al., 2024, Suzuki et al., 25 Dec 2025, Li et al., 7 May 2026, Rahman et al., 29 Jun 2026, Chen et al., 14 Mar 2026).