DarkAgents

Abstract

We present DarkAgents: a multi-agent system that leverages the reasoning and code-generation capabilities of large language models (LLMs), together with deterministic tested human-written code, to build orchestrated pipelines for theoretical astroparticle physics research. While related approaches have been proposed in collider physics and cosmology, DarkAgents targets the specific challenges of this domain, such as model building, complex pipeline computations, multiple constraints and assumption auditing. The framework can be powered by different agentic command-line tools, including Mistral's, Anthropic's, OpenAI's and local LLMs via Ollama. As first implementation, we apply DarkAgents to the study of cosmological first order transitions, starting from a classically scale-invariant particle-physics model and ending with the fit to the NANOGrav nanohertz gravitational-waves spectrum. DarkAgent-PT provides as output i) the best-fit values of model parameters, ii) their existing experimental and observational constraints, iii) an audit report of the assumptions and priors entering both i) and ii), of particular relevance for astroparticle physics. Our test runs identify inconsistencies in some fits in the literature and produce novel ones based on the dissipative bulk-flow GW template. The code is publicly available at https://github.com/PhysicsZandi/DarkAgents.