Mimosa framework enables AI agents to automatically adapt their workflows for scientific research, achieving 43.1% success rate on benchmark tests.

Mimosa is an evolving multi-agent framework that automatically generates and refines task-specific AI workflows through experimental feedback, overcoming limitations of fixed systems

The framework uses Model Context Protocol (MCP) for dynamic tool discovery and includes a meta-orchestrator that designs workflow topologies and code-generating agents

Achieves 43.1% success rate on ScienceAgentBench using DeepSeek-V3.2, outperforming single-agent baselines and static multi-agent configurations

Uses an LLM-based judge to score execution results and provide feedback that drives continuous workflow refinement and improvement