Mimosa-AI
A self-evolving framework that writes, runs and improves its own multi-agent workflows.
Mimosa-AI is an open framework for autonomous scientific research that synthesises task-specific multi-agent workflows and refines them through execution feedback.
The project
Built for research that keeps moving.
Autonomous-research systems are often built on fixed pipelines. Mimosa instead generates, executes and iteratively refines a multi-agent workflow for each scientific task, so the division of labour between agents evolves with the problem.
On ScienceAgentBench (102 data-driven discovery tasks in bioinformatics, computational chemistry, geographic information science and cognitive neuroscience), its iterative-learning mode reached a 43.1% success rate with DeepSeek-V3.2, ahead of single-agent and static multi-agent configurations.
How it works
From a research task to an inspectable workflow.
Tool discovery at runtime
Tools are found through the Model Context Protocol, with the companion platform Toolomics.
Workflow synthesis
A meta-orchestrator assembles a task-specific workflow; code-generating agents run its subtasks.
Learning from execution
A judge scores the outputs and drives the next iteration of the workflow.
Auditable runs
Every workflow topology, agent prompt and execution trace is logged.
Project facts
- Licence
- Apache 2.0
- Team
- Martin Legrand, Tao Jiang, Louis-Félix Nothias
- Partners
- 3iA Côte d’Azur · Wimmics (Inria)