Introduces LEGIT, a credentialing protocol that links certification, reputation, and market mechanisms so buyers can identify which specialized AI agent will perform best on their tasks.
LEGIT is a credentialing protocol designed for trustworthy AI agent marketplaces that addresses the challenge of verifying agent performance by binding measured quality and cost to specific agent configurations, task domains, evaluation budgets, and evidence through a signed record. The protocol operates on three layers: a certification layer that issues capability credentials, a reputation layer that tracks past task outcomes using a discounted Beta update, and a marketplace layer that uses these credentials to facilitate task allocation based on a scoring rule combining quality, cost, reputation, and price. By explicitly recording the model, harness, tools, and parameters used during evaluation, LEGIT ensures that buyers can make informed decisions about which agent configuration offers the best trade-off between performance and resource consumption, while also providing economic incentives to prevent reputation manipulation via Sybil attacks.
LEGIT addresses a core trust problem in AI agent marketplaces: how buyers can determine which specialized agent is likely to perform well on a given task when capabilities, reliability, and task fit are not directly observable. The paper proposes a credentialing protocol that connects three complementary layers—certification, reputation, and market mechanisms—into a single selection framework. Certification supplies baseline, verifiable evidence that an agent has been evaluated against relevant criteria, while reputation captures longitudinal performance signals as the agent is used in practice. Market mechanisms then use these signals to improve discovery, matching, and incentive alignment, making higher-performing agents more competitive and reducing the risk of adverse selection.
The key contribution is a protocol-level design for trustworthy agent procurement rather than a one-off benchmark or trust score. By linking formal credentials to ongoing reputation and market dynamics, LEGIT aims to make agent quality more transparent, auditable, and task-conditioned. This matters because AI agent marketplaces will likely involve heterogeneous, specialized agents with differing safety profiles, capabilities, and failure modes; buyers need defensible ways to compare them without relying solely on vendor claims or a single centralized authority. In that context, LEGIT offers a structured approach to improving accountability, composability, and efficient matching in autonomous agent economies.