Trustworthy Agentic AI: A Governance Framework

Core Idea — Runtime Governance for Agentic AIThis preprint by Clive Aldred proposes a comprehensive governance framework that shifts trustworthy agentic AI from post-execution audits to real-time, cryptographically enforced compliance. It directly addresses the challenges of deploying autonomous AI agents in high-stakes cross-border fintech, crypto settlements, and legal arbitration. Key Components Runtime Agentic Trust Schema (RT-ATS) + Finality Gate: Uses Trusted Execution Environments (TEE) or Multi-Party Computation (MPC) to block non-compliant tool calls before they execute, based on sealed cryptographic predicates. Adaptive Jurisdictional Determinism (A-JD): Combines Retrieval-Augmented Generation (RAG) over legal databases (e.g., case.law) with solver-based reasoning to dynamically determine which jurisdiction’s rules apply and resolve conflicts. Red-Teaming ADTL: An interactive benchmarking suite that detects behavioral drift, prompt injection, and tool poisoning in multi-agent systems using sandboxing and fingerprinting. Dynamic Trust-Utility Frontier (DTUF): Balances trust, utility, uncertainty, differential privacy budgets, and human-in-the-loop costs in real time. Practical RelevanceThe framework is explicitly mapped to the EU AI Act (Articles 9–14), ISO 42001, and MCP security standards. It includes open-source LangGraph/MCP adapters, Merkle-anchored provenance via Agent Identity Documents (AIDs), and W3C Verifiable Credentials for human oversight. The author draws on real-world experience with international payments, Bitcoin Core settlements, regulatory friction, and insurance disputes to argue that verifiable compliance must become infrastructural — not an afterthought. In essence, the paper presents a production-ready “governance substrate” that aims to make autonomous AI agents legally admissible and regulator-ready by design, particularly for fintech and DeFi environments operating across conflicting global rules.