MACC: Multi-Agent Collaborative Competition for Scientific Exploration

A team of researchers from Japan has introduced MACC (Multi-Agent Collaborative Competition), a novel institutional architecture designed to enhance scientific discovery through AI agent collaboration. Published in arXiv and accepted for the AAMAS 2026 Blue Sky Ideas Track, the framework addresses a critical gap in the growing field of MA4Science (Multi-Agent for Science), where most current systems assume all agents are controlled by a single entity. MACC instead examines how institutional mechanisms—like incentives, information sharing, and reproducibility protocols—shape exploration when agents are independently managed, mirroring real-world scientific communities more closely.

The core of MACC is a blackboard-style shared scientific workspace integrated with specific incentive mechanisms to promote transparency, exploration efficiency, and, crucially, reproducibility. This setup provides a controlled testbed to study how different institutional designs influence the reliability and scalability of multi-agent scientific inquiry. The work argues that relying on a single, highly capable LLM agent is insufficient to overcome structural limitations in science, such as redundant trials and limited exploration. By framing AI agents as independent actors within a competitive yet collaborative ecosystem, MACC opens new avenues for designing systems that could accelerate discovery while ensuring robust and verifiable results.