Denario architecture.
These modules are agentic systems that can generate an idea for
a research project, check the literature to validate the novelty
of the idea, develop a plan to carry out the project, implement
the plan by writing and executing code, making plots, and
describing the results, write a scientific paper discussing the
research and showing the plots, and review the paper.
Idea maker-idea hater.
The Idea module consist of two agents: idea maker and idea
hater. The system first calls the idea maker agent, that will
generate an idea given the input text. Next, the idea is sent to
the idea hater, who will critic the generated idea. The idea
maker will then take the input text, the generated idea, and the
critic and will improve the idea. This process is then repeated
several times until the final idea is generated.
Literature.
The literature module evaluates whether an idea is novel by
passing it to a novelty agent, which can classify it as new, not
new, or request more information. If uncertain, the novelty
agent generates queries for the semantic scholar agent, which
retrieves papers until the literature is sufficiently sampled.
If no relevant work is found, the idea is deemed new. Finally, a
summary agent compiles a report explaining the decision and
highlighting key papers.
Planning and control with Cmbagent.
Planning and control implementation of "Deep Research" from
Cmbagent. This is the essential orchestration strategy enabling the
research analysis without human-in-the-loop. Additionally, it
can be used for idea generation, method generation and data
pre-processing.
Denario writing module.
Workflow of the paper writing module. Each node represents an
agent in charge of a task. For instance, the methods node
contains an agent that will write the methods sections
considering the input files.
Denario GUI.
Denario can be run through a GUI called DenarioApp. The user can
input a text prompt and select the modules to run, generating a
full research project from scratch or inputing files for the
different required modules. A demo of this app is hosted on
Hugging Face Spaces.