We're glad to share Mimosa (arXiv, 30 March 2026), an open-source system from researchers at CNRS, Université Côte d'Azur, and Inria that lets artificial-intelligence agents evolve their own strategies for tackling complex scientific problems — and shares everything it does for anyone to check.
The problem: AI systems that can't adapt
Modern science produces enormous amounts of data, but analysing it remains a bottleneck. Most current AI assistants work like a fixed recipe: they follow the same steps regardless of what happens along the way. When an unexpected result appears or a new tool becomes available, they cannot adjust. Real science, however, is messy and non-linear.
The solution: agents that evolve their teamwork
Instead of following a single fixed plan, Mimosa:
- Breaks down a scientific problem and assigns each piece to a specialized AI agent;
- Runs the whole team and evaluates how well it did;
- Learns from the results and reshuffles the team — changing who does what, which tools they use, and how they communicate;
- Repeats the cycle, getting better each time.
Think of it as an AI lab manager that keeps reorganising its research team until it finds the approach that works best for each specific problem.
What makes it special
- It discovers its own tools. Using the Model Context Protocol (MCP) — and the companion platform Toolomics — Mimosa automatically finds and uses new scientific software without human reconfiguration.
- Everything is recorded. Every workflow topology, agent prompt, intermediate file, and execution trace is logged and archived, so any analysis can be replayed step by step.
- It's completely open. Mimosa and Toolomics are released under the Apache 2.0 licence.
The results
On ScienceAgentBench — 102 real scientific tasks drawn from published papers, spanning bioinformatics, computational chemistry, geographic information science, and cognitive neuroscience — Mimosa reached a 43% success rate, beating both a single agent working alone and a fixed multi-agent setup. The gain came specifically from the system's ability to learn and evolve its own workflows over multiple rounds. Mimosa is not locked to any single model.
What the researchers say
"The point is not only to make science faster, but to make more room for the most creative, collective, and responsible parts of scientific work." — Louis-Félix Nothias, CNRS researcher and lead of HolobiomicsLab
"Everything is saved: every workflow topology, every agent prompt, every intermediate file, every memory trace. A major gap is that every single agent decision still relies on attention mechanisms and stochastic sampling, making specific decisions a 'black-box' that is difficult to interpret." — Martin Legrand, CNRS assistant engineer and first author
"When AI manages the execution power of high-throughput experimentation, success depends less on access to armies of technicians and capital equipment, and more on the quality of the initial idea." — Tao Jiang, CNRS research engineer and co-corresponding author
"At Inria we have always supported open science… Trust grows through transparent, auditable codebases countering black-box risks." — Fabien Gandon, Research Director at Inria
This is still early-stage research — Mimosa is a preprint, not yet peer-reviewed — but it is a concrete step toward AI systems that genuinely assist with the complexity of real scientific work. Because it is open-source, the global community can help improve it.
- Preprint: arXiv:2603.28986
- Code: github.com/HolobiomicsLab/Mimosa-AI · Toolomics