EPIAGENT: Agentic AI Systems Overview

• EPIAGENT is a family of agent-based AI systems defined by its use in epidemiological modeling, epidemic planning, clinical prediction, inscription restoration, and other 'epi-' domains.
• It employs structured intermediate representations and verification-first synthesis, ensuring domain-specific constraints and reliable execution of complex simulations and predictions.
• Empirical evaluations show improvements in calibration accuracy, planning efficiency, and detection performance, underscoring the practical impact of iterative refinement in these systems.

Searching arXiv for papers using the term “EPIAGENT” and closely related system names to ground the article in the current literature. EPIAGENT is not a single, universally standardized term in the current literature. Instead, it denotes several distinct computational systems spanning epidemiological modeling, epidemic response planning, clinical prediction, ancient inscription restoration, epigenomic analysis, and EEG interpretation. The term is used most explicitly for an “agentic framework for epidemiological modeling” that synthesizes and validates mechanistic epidemic simulators (Datta et al., 30 Jan 2026), but closely related or variant usages also refer to EpiPlanAgent for automated epidemic response planning (Mao et al., 11 Dec 2025), EpiAgent for inscription restoration (Zhu et al., 10 Apr 2026), and several domain-specific systems or blueprints in which “EPIAGENT” is adopted as a shorthand for an agentic architecture. Accordingly, the most accurate encyclopedic treatment is as a family of agent-based or agentic AI systems rather than a single canonical artifact.

1. Terminological scope and disambiguation

In the epidemiological planning literature, the query term “EPIAGENT” is explicitly aligned with the system name “EpiPlanAgent,” and no alternate “EPIAGENT” naming conventions are reported there (Mao et al., 11 Dec 2025). In a separate epidemiological modeling paper, however, EPIAGENT is the formal system name for an “agentic framework for epidemiological modeling” that automatically synthesizes, calibrates, verifies, and refines simulators (Datta et al., 30 Jan 2026). Other papers use closely related names such as “EpiAgent” for inscription restoration (Zhu et al., 10 Apr 2026) or derive an “EPIAGENT” blueprint from an underlying method, as in rare disease detection for exocrine pancreatic insufficiency (Yu et al., 2019), biological age estimation based on EpiCAge (Jiang et al., 10 Nov 2025), and ARIES-based epidemiological surveillance (Wattamwar et al., 5 Jan 2026).

This naming heterogeneity reflects a broader pattern: “EPIAGENT” functions as a reusable label for agentic systems in domains beginning with “epi-,” including epidemiology, epigraphy, epigenomics, and epilepsy. This suggests that the term should be treated as polysemous. A plausible implication is that any interpretation of EPIAGENT must be anchored to the specific paper and domain in which it appears.

2. EPIAGENT in epidemiological modeling

The most formalized use of EPIAGENT appears in “Agentic Framework for Epidemiological Modeling” (Datta et al., 30 Jan 2026). There, EPIAGENT is an agentic framework that “automatically synthesizes, calibrates, verifies, and refines epidemiological simulators by modeling disease progression as an iterative program synthesis problem” (Datta et al., 30 Jan 2026). Its central abstraction is the Epidemiological Flow Graph intermediate representation, a typed directed multigraph

$G = (V, E, T, \Phi, \Pi, \Omega),$

where nodes encode compartments, parameters, interventions, observations, and source/sink elements, and edges encode epidemiological semantics such as Progression, Transmission, Recovery, Death, Vaccination, Waning, LossOfImmunity, Importation, and ObservationMap (Datta et al., 30 Jan 2026).

The framework is verification-first. It enforces mechanistic invariants before code generation, including non-negativity, dimensional consistency, terminal semantics, and mass conservation in closed systems. In the closed-population case, the paper states that

$\frac{d}{dt}(S+E+I+R)=0$

excluding terminal death compartments (Datta et al., 30 Jan 2026). It also rejects structurally invalid transitions such as $D \to *$, and disallows semantically inconsistent pathways such as direct recovery from susceptibility when the scenario requires latent infection structure (Datta et al., 30 Jan 2026).

Compilation maps verified flow graphs into executable mechanistic simulators, including SEIR-like systems with vaccination and immune escape. One example given is

$$S'(t) = -\beta(t)\frac{S(t)I(t)}{N} - v(t)S(t) + \epsilon R(t),$$

with corresponding equations for $E$, $I$, $R$, and $D$, plus observation models such as

$C_t \sim \mathrm{NB}(\mu_t, k).$

The calibration stage optimizes interpretable epidemiological parameters under constraints and uses a fixed PyTorch skeleton, Adam, and ReduceLROnPlateau (Datta et al., 30 Jan 2026).

Empirically, the paper reports improved calibration accuracy as model flexibility increases. For state-average results, MAE decreases from $3.84\times 10^{-3}$ for time-invariant models to $\frac{d}{dt}(S+E+I+R)=0$0 for time-variant models and $\frac{d}{dt}(S+E+I+R)=0$1 for neural-incorporated models; nationwide MAE decreases from $\frac{d}{dt}(S+E+I+R)=0$2 to $\frac{d}{dt}(S+E+I+R)=0$3 and $\frac{d}{dt}(S+E+I+R)=0$4, respectively (Datta et al., 30 Jan 2026). The ablation results are especially notable: without flow-graph verification, incorrect scenario projections occur in more than 87% of cases, while removal of retrieval augmentation leads to oversimplified graphs and about a 20% increase in graph-generation iterations (Datta et al., 30 Jan 2026).

3. Automated epidemic planning and surveillance variants

A second major epidemiological use of the term refers to EpiPlanAgent, an agent-based system for digital emergency response planning (Mao et al., 11 Dec 2025). The system integrates multi-agent task collaboration, epidemiological knowledge grounding, retrieval augmented generation, structured plan generation, and iterative refinement (Mao et al., 11 Dec 2025). Its workflow is built on SigmaFlow, which is formalized as a directed graph with node set

$\frac{d}{dt}(S+E+I+R)=0$5

comprising Model Nodes, Tool Nodes, and Logic Nodes (Mao et al., 11 Dec 2025).

EpiPlanAgent covers 8 infectious diseases and uses a structured knowledge base drawn from “over 40 textual data resources” from official public health guidance (Mao et al., 11 Dec 2025). Each knowledge-base entry is an atomic JSON action with fields including Action, Trigger Condition, Work Requirement, Responsible Party, Time Limit, and Termination Condition (Mao et al., 11 Dec 2025). The system uses exact-match RAG rather than vector retrieval in order to avoid semantic drift and preserve guideline alignment (Mao et al., 11 Dec 2025).

The evaluation involved $\frac{d}{dt}(S+E+I+R)=0$6 public health professionals and $\frac{d}{dt}(S+E+I+R)=0$7 real-world-inspired scenarios (Mao et al., 11 Dec 2025). Plan completeness improved from $\frac{d}{dt}(S+E+I+R)=0$8 for manual planning to $\frac{d}{dt}(S+E+I+R)=0$9 in Round 1 and $D \to *$0 in Round 2, with statistical significance against manual planning of $D \to *$1 and $D \to *$2, respectively (Mao et al., 11 Dec 2025). Plan generation time decreased from $D \to *$3 minutes manually to $D \to *$4 minutes in Round 1 and $D \to *$5 minutes in Round 2; the main reported reduction is 93.9% (Mao et al., 11 Dec 2025). Alignment between AI-generated and expert-authored section scores reached Pearson $D \to *$6 with 95% CI $D \to *$7–$D \to *$8, $D \to *$9, and 91.5% of participants rated the system “very helpful” or “extremely helpful” for efficiency and standardization (Mao et al., 11 Dec 2025).

A related but distinct system is ARIES, a hierarchical multi-agent framework for epidemiological surveillance and outbreak monitoring (Wattamwar et al., 5 Jan 2026). ARIES uses a Manager agent plus domain-specific sub-agents for WHO Disease Outbreak News, CDC WONDER, and PubMed, implemented in Python with CrewAI (Wattamwar et al., 5 Jan 2026). Its reported evaluation emphasizes output breadth and citation coverage rather than formal detection metrics. In a tested Mpox Clade Ib scenario, the configuration using a gpt-5.1 Manager with o3 sub-agents produced a 2,962-word report with 20 sources, outperforming shallower configurations in source coverage and detail (Wattamwar et al., 5 Jan 2026). No precision, recall, F1, AUROC, or time-to-detect statistics are reported (Wattamwar et al., 5 Jan 2026).

These systems collectively define a recognizable epidemiological EPIAGENT pattern: task decomposition, authoritative retrieval, structured intermediate artifacts, explicit validation, and strong emphasis on provenance.

4. Clinical and biomedical prediction uses

Outside epidemiology proper, EPIAGENT is also used as a blueprint or system label in biomedical prediction. In “Rare Disease Detection by Sequence Modeling with Generative Adversarial Networks” (Yu et al., 2019), EPIAGENT refers to a sequence-modeling system for early detection of exocrine pancreatic insufficiency from longitudinal claims data. The system combines a single-layer LSTM with a semi-supervised GAN operating on learned patient embeddings (Yu et al., 2019). The longitudinal input is encoded using 300-dimensional medical-code embeddings, sequences are padded or truncated to $$S'(t) = -\beta(t)\frac{S(t)I(t)}{N} - v(t)S(t) + \epsilon R(t),$$0, and the LSTM generates a $$S'(t) = -\beta(t)\frac{S(t)I(t)}{N} - v(t)S(t) + \epsilon R(t),$$1 representation; after appending age and gender, the final feature vector has dimension 258 (Yu et al., 2019).

The discriminator is a $$S'(t) = -\beta(t)\frac{S(t)I(t)}{N} - v(t)S(t) + \epsilon R(t),$$2-way classifier with $$S'(t) = -\beta(t)\frac{S(t)I(t)}{N} - v(t)S(t) + \epsilon R(t),$$3 supervised classes plus a fake class, and the model uses labeled, unlabeled, generated, entropy, feature-matching, and pull-away losses (Yu et al., 2019). On a test set with prevalence approximately 1.30%, the SSL-GAN achieved PR-AUC $$S'(t) = -\beta(t)\frac{S(t)I(t)}{N} - v(t)S(t) + \epsilon R(t),$$4, compared with $$S'(t) = -\beta(t)\frac{S(t)I(t)}{N} - v(t)S(t) + \epsilon R(t),$$5 for the best supervised DNN baseline, described as a 6% relative improvement (Yu et al., 2019). The dataset spans 7 years of claims from approximately 1.8 million patients, including 29,149 labeled EPI cases (Yu et al., 2019).

A different biomedical instantiation appears in “Integrating Epigenetic and Phenotypic Features for Biological Age Estimation in Cancer Patients via Multimodal Learning” (Jiang et al., 10 Nov 2025). There, EpiCAge is the underlying model, but the paper explicitly describes “EPIAGENT based on EpiCAge” as a clinical decision-support framework (Jiang et al., 10 Nov 2025). EpiCAge is a two-layer stacked multimodal ensemble combining an epigenetic clock, a phenotypic clock, and an early-fusion clock, followed by late fusion with a PCA-based skip connection (Jiang et al., 10 Nov 2025). The estimator is written as

$$S'(t) = -\beta(t)\frac{S(t)I(t)}{N} - v(t)S(t) + \epsilon R(t),$$6

with age acceleration defined by

$$S'(t) = -\beta(t)\frac{S(t)I(t)}{N} - v(t)S(t) + \epsilon R(t),$$7

Across eight internal cohorts, EpiCAge-TabPFN achieved RMSE $$S'(t) = -\beta(t)\frac{S(t)I(t)}{N} - v(t)S(t) + \epsilon R(t),$$8, MAE $$S'(t) = -\beta(t)\frac{S(t)I(t)}{N} - v(t)S(t) + \epsilon R(t),$$9, and $E$0; across four external cohorts, it achieved RMSE $E$1, MAE $E$2, and $E$3, outperforming listed epigenetic and phenotypic baselines (Jiang et al., 10 Nov 2025). Age acceleration was significantly associated with mortality, with hazard ratio $E$4 internally and $E$5 externally per 5-year increase, both $E$6 (Jiang et al., 10 Nov 2025).

These biomedical uses differ from the epidemiological systems in their architectures but retain a common theme: EPIAGENT denotes an operational layer that turns predictive models into structured, decision-oriented workflows.

5. Extension to other agentic domains

The term also appears in non-biomedical contexts. “EpiAgent: An Agent-Centric System for Ancient Inscription Restoration” (Zhu et al., 10 Apr 2026) formulates restoration as hierarchical planning over multimodal evidence, coordinated by a central LLM planner in an Observe–Conceive–Execute–Reevaluate loop (Zhu et al., 10 Apr 2026). The observation record is

$E$7

and per-character plans are

$E$8

Reevaluation is based on text authenticity $E$9, style consistency $I$0, and optional human feedback (Zhu et al., 10 Apr 2026).

On the synthetic set $I$1, EpiAgent achieved PSNR $I$2, SSIM $I$3, LPIPS $I$4, CLIP-IQA $I$5, MUSIQ $I$6, MANIQA $I$7, NIMA $I$8, Top-1 accuracy $I$9, Top-5 accuracy $R$0, Macro accuracy $R$1, and 1-NED $R$2, with best results across all reported metrics relative to the named baselines (Zhu et al., 10 Apr 2026). A user study with more than 20 experts reported Top-1 preference 59.66%, Top-3 84.18%, and mean ranking 82.11% (Zhu et al., 10 Apr 2026).

“EEG-SpikeAgent: Agentic Closed-Loop Program Synthesis for Automated EEG Spike Detection” (Santhosh et al., 6 Jul 2026) provides another agentic variant. It uses an LLM to propose deterministic EEG feature modules, which are then executed, evaluated by XGBoost, and iteratively refined (Santhosh et al., 6 Jul 2026). On VEPISET, the resulting system achieved AUROC $R$3, balanced accuracy $R$4, F1 $R$5, sensitivity $R$6, and specificity $R$7 at the default threshold; at sensitivity $R$8, mean precision was $R$9 and mean specificity $D$0 (Santhosh et al., 6 Jul 2026). Artifact-aware feature generation improved balanced accuracy and F1 relative to spike-only feature search (Santhosh et al., 6 Jul 2026).

These uses make clear that EPIAGENT, in contemporary arXiv practice, designates a general agentic pattern rather than a single domain-locked architecture.

6. Shared architectural motifs

Despite domain differences, several recurrent design choices unify EPIAGENT-like systems.

Motif	Reported form	Example
Explicit intermediate structure	Flow graph, JSON task list, feature manifest, per-character plan	(Datta et al., 30 Jan 2026, Mao et al., 11 Dec 2025, Santhosh et al., 6 Jul 2026, Zhu et al., 10 Apr 2026)
Verification before or during execution	Static graph checks, exact-match RAG grounding, reevaluation metrics, deterministic grading	(Datta et al., 30 Jan 2026, Mao et al., 11 Dec 2025, Zhu et al., 10 Apr 2026, Muralidharan et al., 11 Jun 2026)
Specialized tool use	CDC WONDER, WHO DONs, PubMed, diffusion tools, code execution, OCR, XGBoost	(Wattamwar et al., 5 Jan 2026, Zhu et al., 10 Apr 2026, Santhosh et al., 6 Jul 2026)
Iterative refinement	Feedback rounds, replanning, agent loop diagnostics	(Mao et al., 11 Dec 2025, Zhu et al., 10 Apr 2026, Santhosh et al., 6 Jul 2026, Datta et al., 30 Jan 2026)
Domain-constrained outputs	Strict JSON, fixed simulator skeleton, deterministically gradable answers	(Mao et al., 11 Dec 2025, Datta et al., 30 Jan 2026, Muralidharan et al., 11 Jun 2026)

The benchmark paper “EpiBench: Verifiable Evaluation of AI Agents on Epigenomics Analysis” (Muralidharan et al., 11 Jun 2026) is particularly relevant as a meta-level statement about such systems. EpiBench evaluates short-horizon epigenomics decisions using deterministic graders across 106 evaluations and 5,088 valid trajectories (Muralidharan et al., 11 Jun 2026). No model-harness pair passed a majority of attempts; the best endpoint pass rate was 45.0% for GPT-5.5 / Pi, with 95% CI 36.3–53.7 (Muralidharan et al., 11 Jun 2026). Field-level scoring was 68.2%, whereas endpoint pass rate was 31.0%, indicating that agents often compute correct intermediate results but fail at assay-specific judgment (Muralidharan et al., 11 Jun 2026).

This benchmark result suggests that the core difficulty for EPIAGENT-like systems is not mere tool invocation. Rather, the decisive bottleneck is reliable domain judgment under structured constraints.

7. Limitations, misconceptions, and research trajectory

A common misconception would be to treat EPIAGENT as a single mature platform. The literature does not support that interpretation. Instead, different papers use the label for different systems, with differing levels of maturity, reproducibility, and evaluation. For example, EpiPlanAgent reports controlled user evaluation with public health professionals but does not report full code or knowledge-base release (Mao et al., 11 Dec 2025). ARIES provides architectural descriptions and scenario outputs but not formal outbreak-detection benchmarks (Wattamwar et al., 5 Jan 2026). The epidemiological modeling EPIAGENT offers formal guarantees and ablations but focuses on scenario case studies rather than real-time operational deployment (Datta et al., 30 Jan 2026).

Another misconception would be that “agentic” automatically implies reliability. Several papers emphasize safeguards precisely because LLM reasoning is error-prone. EpiPlanAgent relies on authoritative grounding, exact-match RAG, strict JSON schema, and expert panel evaluation (Mao et al., 11 Dec 2025). The modeling EPIAGENT prevents degeneration by verification-first synthesis and rejects structurally invalid models rather than masking errors by parameter clipping (Datta et al., 30 Jan 2026). EpiBench shows that even strong model-harness pairs fail frequently when domain-specific scientific judgment is required (Muralidharan et al., 11 Jun 2026).

Future directions are reported with notable consistency across the epidemiological papers: dynamic knowledge integration, cross-jurisdictional adaptation, multi-agent simulation and verification, and more explicit human-in-the-loop review for public health planning (Mao et al., 11 Dec 2025); Model Context Protocol, RAG, and additional data streams for surveillance (Wattamwar et al., 5 Jan 2026); and online data assimilation, extended heterogeneity, and interpretable neural residuals for epidemiological modeling (Datta et al., 30 Jan 2026). This suggests a convergent research agenda in which EPIAGENT evolves toward modular, tool-grounded systems with stronger validation layers and tighter coupling between structured representations and domain-specific constraints.

In sum, EPIAGENT is best understood as a class of agentic, domain-constrained AI systems that elevate intermediate representations, verification, and iterative refinement to first-class design principles. Its most rigorous instantiation to date is the epidemiological modeling framework built around the Epidemiological Flow Graph IR (Datta et al., 30 Jan 2026), but the term now spans a broader landscape of task-specific agent systems whose common ambition is to replace brittle single-pass inference with auditable, structured, and evaluable workflows.