LEGIT Proposes Verifiable Credentials for AI Agent Marketplaces

Summary

AI agent marketplaces are emerging, but buyers may struggle to determine which agent will perform best because benchmark scores can be difficult to verify or compare across tasks, software, and budgets. The paper introduces LEGIT, a credentialing protocol that connects certification, reputation, and marketplace allocation. Its certification records bind measured quality and cost per solved task to a specific agent configuration, task domain, evaluation budget, and supporting evidence, with the record signed for verification. Reputation links the outcomes of past tasks to the same identity, while reliability depends on the feedback being reported. Buyers and agents can verify credential records and inspect optional visual profiles. The reported evaluations find that agent configurations with similar observed success can have different costs, and that comparisons change with the evaluation budget. These findings motivate tying performance claims to the exact configuration and resource limits used in testing. A complementary analysis estimates the deposits and fees needed to manipulate reputation under a stated Sybil attack model.