SAGA: Shapley-Governed Agentic Governance Architecture for Multi-Enterprise AI Coalitions
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Autonomous AI agents working across enterprise boundaries today carry governance as an afterthought: access-control policies that are attached externally, can be circumvented, and give no neutral auditable record that all parties trust. This paper presents SAGA (Shapley-Governed Agentic Governance Architecture), a framework that embeds governance directly into agent identity by encoding each agent's Shapley-derived contribution score, permission tier, and coalition memberships into a cryptographically signed agent capability descriptor. SAGA rests on four contributions: governance-intrinsic agent identity extending the Google A2A and SAP ORD card schemas; a two-layer hierarchical Shapley governance system in which intra-enterprise LLM agent scores feed cross-enterprise org-agent characteristic functions; event-triggered self-healing governance where a violation atomically triggers Shapley recomputation, tier downgrade, coalition revocation, and card re-issuance in a single on-chain transaction; and a hybrid blockchain design combining Hyperledger Fabric for permissioned operational state with periodic Merkle-root anchoring to Ethereum for regulatory auditability, connected by a ZKP commitment scheme. A complete implementation is provided with 74 Python, 5 Go, and 4 Solidity tests all passing. Experiments report Monte Carlo approximation error converging at O(1/sqrt(k)), linear-time Shapley computation scaling, sub-millisecond ZKP commit-verify throughput, and quantitative violation-recovery dynamics across all governance tiers.