active · HolobiomicsLab · 3iA Côte d'Azur · Wimmics (Inria)
Mimosa — evolving multi-agent systems for science
A self-evolving AI framework for autonomous scientific research (ASR) that writes, runs, and improves its own multi-agent workflows — powered by MCP tool discovery and Darwinian evolution.
Mimosa is the lab's flagship multi-agent AI system — a self-evolving framework for Autonomous Scientific Research (ASR) that writes, runs, and improves its own multi-agent workflows, powered by MCP tool discovery and Darwinian evolution. Unlike autonomous-research systems built on fixed pipelines, Mimosa automatically generates, executes, and iteratively refines task-specific multi-agent workflows — letting the system evolve its own division of labour for each scientific task.
On ScienceAgentBench (102 data-driven discovery tasks across bioinformatics, computational chemistry, GIS, and cognitive neuroscience), Mimosa's iterative-learning mode reached a 43.1% success rate with DeepSeek-V3.2 — surpassing both single-agent and static multi-agent configurations. It uses the Model Context Protocol (MCP) for runtime tool discovery, and logs every workflow topology, agent prompt, and execution trace for full auditability.
Released under the Apache License 2.0 alongside the companion platform Toolomics. First-authored by Martin Legrand.
- Preprint: arXiv:2603.28986 (30 March 2026)
- Code: github.com/HolobiomicsLab/Mimosa-AI