Denario: AI-Driven Multi-Agent Discovery

• Denario is an open-source, modular multi-agent system that decomposes scientific research into specialized agent roles for rapid and autonomous discovery.
• It employs asynchronous message-passing and advanced algorithms for literature synthesis, hypothesis generation, model criticism, and verification.
• Benchmark evaluations show Denario significantly improves discovery rates and reduces compute costs compared to human-only approaches.

Denario is an open-source, modular multi-agent system for AI-driven scientific discovery, designed for autonomous or semi-autonomous generation, analysis, and critique of scientific research. Denario decomposes the research process into specialized agent roles coordinating via asynchronous message-passing, enabling rapid literature synthesis, hypothesis generation, model criticism, verification, and manuscript drafting. Its architecture, cross-domain capabilities, and design of epistemic processes position it as both a practical research assistant and a reference implementation of the “AI scientist” concept (Jimenez et al., 22 Jun 2026, Villaescusa-Navarro et al., 30 Oct 2025).

1. System Architecture and Agent Roles

Denario operates as a circle of modular agents—each focused on a classical research function—coordinated by an orchestration layer. Roles are strictly partitioned, with agents communicating over a message bus to maintain causal traceability and high-throughput parallelism. The canonical roles are:

• Literature Synthesis Agent (“Retriever”): Continuously mines archives (arXiv, ADS, PubMed), builds a structured knowledge graph, and provides entity-based summaries.
• Hypothesis Generation Agent (“Generator”): Consumes literature summaries to propose candidate models or theoretical extensions, using combinatorial recombination and language-model-driven extrapolation.
• Model Criticism Agent (“Critic”): Evaluates proposed hypotheses for internal consistency, constraint violations, and penalizes over-parameterization.
• Verification Agent (“Verifier”): Instantiates candidate models into executable code for empirical or statistical validation, producing fitness metrics (e.g., log-likelihood, Bayes factor, χ²).
• Controller Agent (“Coordinator”): Triggers retrieval, generation, criticism, and verification stages, maintaining stateful memory across the discovery cycle.

Denario’s modular architecture enables agents to be invoked independently (e.g., only Analysis) or as a complete pipeline. Integration with “Cmbagent” provides a deep-research backend for planning, debugging, subtask decomposition, and robust code execution (Villaescusa-Navarro et al., 30 Oct 2025).

2. Core Algorithms and Workflow Formalism

Denario implements reproducible, multi-stage workflows for hypothesis proposal, critical evaluation, and end-to-end paper drafting. Communication is formalized as:

$$M := \langle \text{sender\_id}, \text{receiver\_id}, \text{task\_type}, \text{payload} \rangle$$

with payloads supporting text, LaTeX equations, code, JSON, or evaluation scores. The Controller formalizes state updates as

$$\text{state}_{t+1} = \text{Controller}(\text{state}_t, M_{\text{Critic}\to\text{Controller}}, M_{\text{Verifier}\to\text{Controller}}, \ldots)$$

Hypothesis proposal proceeds in cycles: retrieve context, generate hypotheses under constraints, check for violations, evaluate surviving candidates, and log both advances and rejections.

Mathematical objective functions:

• Critic cost:

$$C(H) = \lambda_1 \sum_{v \in V(H)} w_v + \lambda_2\,\text{Complexity}(H)$$

• Verifier reward:

$$R(H|D) = \log p(D|H) - \alpha\,\text{Complexity}(H)$$

• Selection:

$$H^* = \arg\max_H [R(H|D) + \beta\,\text{Novelty}(H)]$$

Novelty is estimated via a distance metric in model embedding space.

Analysis leverages Plan–Control loops with engineer and researcher agents: engineers write and execute Python code, generate plots, and return complete results; researchers produce in-depth results sections for manuscript integration. The control agent handles error correction and ensures progression within capped iterations (Villaescusa-Navarro et al., 30 Oct 2025).

3. Empirical Evaluation and Quantitative Metrics

Denario’s performance has been benchmarked in domains such as modified gravity (DHOST) model recovery and cross-disciplinary paper generation. The following summarizes results from (Jimenez et al., 22 Jun 2026):

Method	Discovery Rate	False-Positive Rate	Compute Cost (GPUh)	Novelty Score
Human-only	0.20	0.05	200	—
Denario (v1)	0.65	0.12	50	0.32
Denario (v2*)	0.82	0.08	80	0.45

*v2 includes improved symbolic-solver integration.

The discovery rate for ghost-free DHOST models increases by a factor of ≳3–4 compared to human-only search, with a moderate rise in false positives and a fivefold reduction in compute cost. A typical workflow can generate and evaluate 1000 symbolic model proposals with dynamic pruning through Critic and Verifier loops, enabling systematic exploration of 10⁶–10⁸ model topologies within hours.

Table 1 from (Villaescusa-Navarro et al., 30 Oct 2025) details domain expert evaluation on 10 material-science prompts:

Metric	Range	Mean	Success %	Notes
Context Understanding	0–3	3.00	100%	Problem ID perfect
Method Selection	0–3	2.25	75%	MSD, RDF, density
Scientific Insights	0–3	1.37	45%	Qualitative only
Novel Insights	0–3	1.00	33%	No advanced

Papers generated by Denario are rated with a mean of ≈6.8/10 by domain referees, with human experts noting perfect problem identification and methodological adequacy, but limited depth of interpretation and novel insight, comparable to a senior graduate student (Villaescusa-Navarro et al., 30 Oct 2025).

4. Cross-Disciplinary Capabilities and Use Cases

Denario demonstrates generality by supporting astrophysics, mathematics, biology, medicine, chemistry, neuroscience, and planetary science. Notable workflows include:

• Generation of observational pipelines for gravitational-wave data (dimensionality reduction, cluster analysis, divergence quantification).
• Symbolic regression for mathematical physics, including PINN-based manifold exploration of PDEs.
• Novel cross-field synthesis, e.g., applying quantum tensor train (QTT) compression to astrophysical graph structures.

Output artifacts are standards-compliant: manuscript drafts (LaTeX + BibTeX), plots, code (Python), and review reports (PNG-marked referee comments). Module usage is individualized or chained for complete, reproducible pipelines; all intermediate artifacts are checkpointed for auditability.

5. Institutional, Epistemic, and Ethical Considerations

AI-driven research systems such as Denario introduce requirements for scientific institution redesign to address provenance, accountability, and interpretability (Jimenez et al., 22 Jun 2026):

• Verification: Machine-readable provenance for agentic decisions (timestamps, versions, data lineage) is enforced.
• Accountability: Human–AI interactions are auditable; human sign-off is required for any publication draft.
• Interpretability: Outputs must be accompanied by natural-LLM explanations and documentation of salient features or failure modes.

Regarding authorship, Denario agents do not merit authorship; responsible authorship resides with human supervisors who interpret and contextualize outputs. Peer review incorporates automated Verifier checks as front-end filters, with final judgments on novelty and real-world impact by human referees.

Potential epistemic risks include the entrenchment of “normal science,” citation gaps, and susceptibility to undetected LLM errors and hallucinations. The system’s environmental cost is non-trivial, considering LLM inference and fine-tuning (Villaescusa-Navarro et al., 30 Oct 2025). Malicious usages—such as automated generation of fake research or citation manipulation—are recognized as vectors for harm, necessitating robust institutional safeguards.

Conceptual connections are drawn to Kuhn’s paradigms, Lakatos’s research programmes, Quine's web of belief, and Feyerabend’s epistemic anarchism, positioning Denario as both a disruptor and a potential homogenizer of scientific workflows. Human-AI collaboration, rather than “Turing Trap” mimicry, is recommended to maximize epistemic benefit and creativity.

6. Implementation, Reproducibility, and Accessibility

Denario is fully open source:

• Core repository: https://github.com/AstroPilot-AI/Denario
• Deep-research backend (Cmbagent): https://github.com/CMBAgents/cmbagent
• Orchestration (AG2): https://github.com/ag2ai/ag2
• Web demo: https://huggingface.co/spaces/astropilot-ai/Denario

Installation (Python):

pip install denario[all]

End-to-end workflow invocation:

from denario import Denario, Journal den = Denario(project_dir="myproj") den.set_data_description(input_text) den.get_idea() den.check_idea() den.get_method() den.get_results() den.get_paper(journal=Journal.APS) den.referee()

The browser-based UI offers module selection, data upload, LLM tuning, and batch download. All example papers, data, and prompt templates are provided for direct reproducibility.

This demonstrates that Denario provides a modular, transparent, and scalable foundation for AI-assisted research, with performance rapidly approaching—and in throughput, surpassing—expert human workflows, though ultimate accountability, interpretation, and epistemic validation remain human responsibilities (Jimenez et al., 22 Jun 2026, Villaescusa-Navarro et al., 30 Oct 2025).