Researchers have introduced Mimosa, an open-source framework designed to automate scientific research through evolving multi-agent systems that adapt their workflows based on experimental feedback, according to arxiv.org.
According to the paper published on April 1, 2026, Mimosa addresses limitations in current Autonomous Scientific Research (ASR) systems, which “remain constrained by fixed workflows and toolsets that prevent adaptation to evolving tasks and environments.” The framework automatically synthesizes task-specific multi-agent workflows and iteratively refines them through feedback.
The system leverages the Model Context Protocol (MCP) for dynamic tool discovery and uses a meta-orchestrator to generate workflow topologies. Code-generating agents execute subtasks by invoking available tools and scientific software libraries, while an LLM-based judge scores executions and drives workflow refinement, according to arxiv.org.
On ScienceAgentBench testing, Mimosa achieved a 43.1% success rate using DeepSeek-V3.2, “surpassing both single-agent baselines and static multi-agent configurations,” the paper states. The researchers note that “models respond heterogeneously to multi-agent decomposition and iterative learning,” indicating benefits depend on the underlying execution model’s capabilities.
According to arxiv.org, the framework’s “modular architecture and tool-agnostic design make it readily extensible,” while fully logged execution traces support auditability. The 48-page paper was released as a fully open-source platform to provide “an open foundation for community-driven ASR.”