Reproduce “LOGOS: A Living Logic for AI Agent Teams That Evolve With Humans” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yuma Ichikawa, Yamato Arai, Kosaku Kimura, Akira Sakai, Hiromichi Kobashi

Researchers from Fujitsu Limited, RIKEN Center for AIP, and The University of Tokyo introduce LOGOS, a governance layer for self-evolving AI agent teams that controls adaptation and ensures accountability. It provides an evidence-governed release contract, allowing agents to propose changes while human-owned policies authorize their promotion, demonstrating effective mitigation of 'misevolution' and unauthorized actions in simulated environments.

Key insights

  • The core principle 'proposal is not promotion' is essential for governing self-evolving agents, requiring proposed changes to pass independently defined evidence and human-controlled policy before deployment.
  • A unified governance contract for artifacts, events, verification, authority, promotion, and audit enables pluggable self-evolution across diverse multi-agent systems, integrating concerns often treated separately.
  • Implementing a verifiable human-agent loop, where agents can propose improvements and ask for clarification, while humans steer via explicit approvals and root policies, transforms collaboration into an observable, testable, and governable feedback mechanism.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.