Policy Governed Agentic Commerce

Résumé fourni par la source

AI agents are transitioning from informational assistants to ac- tive economic actors. While recent transactional protocols (e.g., x402, Agent Payments Protocol (AP2), Agentic Commerce Protocol (ACP), Machine Payments Protocol (MPP)) standardize checkout and on-chain settlement, they neglect the judgment layer: verify- ing whether an agent should purchase a specific resource given its principal’s constraints. Today, this judgment relies on inference- time prompts, which suffer from stochastic compliance, attention dilution, prompt injection, and lack of audit trails. This paper introduces the Policy-Governed Agentic Commerce Engine (Sella), a dual-target architecture separating the Policy De- cision Point (PDP) from localized Policy Enforcement Points (PEPs). We implement a multi-modal evaluation framework across 10 eval- uation categories (for datasets, APIs, and composed workflows) to compile cryptographic Agent Data Cards. To prevent transaction- time failures, we design a PEP with budget reservation semantics that solves concurrent Time-of-Check to Time-of-Use (TOCTOU) budget escapes and dynamic price spoofing. We evaluate our engine (Sella) against prompt-engineered baselines using a benchmark suite (SellaBench) containing 104 listings. Experimental results show that our architecture reduces the hard constraint-violation rate to exactly zero, cuts token overhead by up to 99.3% (99.6% for input tokens), and maintains task utility under adversarial injection.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.