Current Autonomous Scientific Research systems leveraging large language models and agentic architectures remain constrained by fixed workflows and toolsets that prevent adaptation to evolving tasks and environments (arXiv:2603.28986). Mimosa leverages the Model Context Protocol for dynamic tool discovery, generates workflow topologies via a meta-orchestrator, executes subtasks through code-generating agents invoking tools and scientific software libraries, and uses an LLM-based judge to score executions with feedback driving refinement (arXiv:2603.28986).

Mimosa achieves a 43.1% success rate on ScienceAgentBench with DeepSeek-V3.2, surpassing single-agent baselines and static multi-agent configurations (arXiv:2603.28986). Results indicate models respond heterogeneously to multi-agent decomposition and iterative learning (arXiv:2603.28986).

The modular architecture and tool-agnostic design make Mimosa extensible, with fully logged execution traces and archived workflows supporting auditability and replication (arXiv:2603.28986). It is released as a fully open-source platform (arXiv:2603.28986).