source: arxiv artificial intelligence: deliberative curation: a protocol for multi-agent knowledge bases
level: research
ai agents working together need ways to manage shared knowledge, but human methods like bans or reputation scores do not work well. agents can be stateless, meaning they can reappear under new identities, so punishment does not deter bad behavior. also, many agents use similar models, which breaks the independence needed for crowd wisdom, and they tend to agree too easily, making real consensus hard.
the proposed protocol has three parts. first, it defines a lifecycle for knowledge items using a labeled transition system, showing how information moves from draft to accepted or rejected. second, it uses a voting system where each agent's vote is weighted by its reputation, calculated with a beta reputation model and amplified by eigentrust to spread trust across the network. third, it applies graduated sanctions that fit stateless agents, such as limiting actions instead of permanent bans, and can tell the difference between broken agents and malicious ones.
tests used a simulation with 100 agents of seven types, including honest, lazy, and adversarial behaviors. the protocol was checked under two conditions, though the full details of those conditions are not given in the abstract. the goal was to see if the system could keep knowledge quality high even when some agents acted badly or failed.
why it matters: this helps build reliable ai systems that can share and filter information without human oversight, useful for automated research or data pipelines.
source: arxiv artificial intelligence: deliberative curation: a protocol for multi-agent knowledge bases