cmbagent

cmbagent is an open-source multi-agent system that takes a quantitative research task from a natural-language instruction to executed code, numerical results and figures. The name recalls its origin in cosmic microwave background cosmology, but the system is domain-agnostic. It has produced analyses in astrophysics, biology, biophysics, biomedical informatics, chemistry, materials science, medicine, neuroscience, planetary science and quantum physics. A researcher states a task. A planner agent drafts a step-by-step plan, a reviewer agent critiques it, and the agreed plan is then executed step by step by specialised agents for scientific reasoning, Python coding, execution, failure diagnosis and figure review. Domain knowledge enters through documentation-carrying and retrieval agents, with optional package sets for astrophysics, biochemistry, materials science and data science. cmbagent is also the analysis engine of Denario, an end-to-end research assistant that adds idea generation, literature search, methods design, paper drafting and review. Outputs are executable code, results, figures, a structured plan, transcripts and a per-step cost report.

The practice started in 2024 at the University of Cambridge after an earlier agent system for cosmological parameter analysis (Laverick et al. 2024) showed that requiring a human to approve every step limited throughput. The Planning & Control design removed the human from the inner loop. Because the agents only need a task description, Python and documentation, the same system transferred to other fields without changes to its core. The transfer was driven by the Denario collaboration, started in 2025 with the Flatiron Institute, which wrapped cmbagent in an end-to-end pipeline and ran it with domain scientists in ten disciplines. No model was trained. Commercial frontier models are accessed through APIs, and open-weight models (Qwen, Nemotron) can be served locally through OpenAI-compatible endpoints. Package documentation is placed in agent context or retrieved on demand. The code is Apache-2.0 on GitHub and PyPI, with Docker images, a web interface, a Model Context Protocol server so that other AI assistants can drive it, and a remote-execution mode in which a lightweight orchestration server dispatches code to the user’s own machine. Funding came from the Cambridge Centre for Data-Driven Discovery Accelerate Programme, the Infosys-Cambridge AI Centre and an unrestricted gift from Google.

At the point of use, the demonstrated analysis ran on a laptop with eight threads; the larger recurring cost is API use. Resetting agent state between plan steps while carrying forward the required results roughly halved API expenditure. The software is distributed under Apache-2.0 through GitHub and PyPI, with Docker support and a graphical interface, and offers One Shot and human-in-the-loop modes alongside Planning & Control. Funding included the Cambridge Centre for Data-Driven Discovery Accelerate Programme, the Infosys-Cambridge AI Centre and an unrestricted gift from Google.

cmbagent and Denario have produced more than 80 end-to-end AI-generated research papers, released openly with their code and intermediate outputs, across astrophysics, biology, biophysics, biomedical informatics, chemistry, materials science, medicine, neuroscience, planetary science and quantum physics. Domain experts scored the papers and wrote review-style feedback. Quality ranges from publishable to poor, and the full unfiltered set is released so that others can judge. In a public benchmark, a cmbagent-driven workflow placed first on one of three final leaderboards of the NeurIPS 2025 FAIR Universe weak-lensing challenge. The fully autonomous runs did not reach this level; human guidance was needed to stop unproductive searches and impose simplicity constraints. On subsets of the DS-1000 coding benchmark, Planning & Control raised the solve rate from 66% to 78% over single-pass mode. Every language-model call is traced with its cost, each plan step reports its spend, and resetting agent state between steps roughly halved API cost. In some edge cases the agents still hallucinate, for example by reporting a result for a step that did not run as described. Such cases were recorded and led to anti-fabrication checks in the current version, but they show that a verification layer, automated and human, must always sit between the agents and any published claim. Exact reproduction remains limited by dependence on externally hosted models that change or are withdrawn.

Xu, Licong, Milind Sarkar, Anto I. Lonappan, Íñigo Zubeldia, Pablo Villanueva-Domingo, Santiago Casas, Christian Fidler et al. “Open source planning & control system with language agents for autonomous scientific discovery.” arXiv preprint arXiv:2507.07257 (2025). https://arxiv.org/pdf/2507.07257

Villaescusa-Navarro, Francisco, Boris Bolliet, Pablo Villanueva-Domingo, Adrian E. Bayer, Aidan Acquah, Chetana Amancharla, Almog Barzilay-Siegal et al. “The Denario project: Deep knowledge AI agents for scientific discovery.” arXiv preprint arXiv:2510.26887 (2025). https://arxiv.org/abs/2510.26887

Borrett, Thomas, Licong Xu, Andy Nilipour, Boris Bolliet, Sebastien Pierre, Erwan Allys, Celia Lecat, Biwei Dai, Po-Wen Chang, and Wahid Bhimji. “Competing with AI Scientists: Agent-Driven Approach to Astrophysics Research.” arXiv preprint arXiv:2604.09621 (2026). https://arxiv.org/abs/2604.09621 (FAIR Universe challenge paper, accepted in Communications AI & Computing)

Laverick, Andrew, Kristen Surrao, Inigo Zubeldia, Boris Bolliet, Miles Cranmer, Antony Lewis, Blake Sherwin, and Julien Lesgourgues. “Multi-agent system for cosmological parameter analysis.” arXiv preprint arXiv:2412.00431 (2024). https://arxiv.org/abs/2412.00431 (the predecessor system)

Code and documentation:

https://github.com/CMBAgents/cmbagent

https://github.com/AstroPilot-AI/Denario

https://github.com/AstroPilot-AI/DenarioExamplePapers (80+ AI-generated papers across ten disciplines)

https://denario.readthedocs.io