AEGIS: Agent Execution Governance and Integrity Standard — A Survey and Reference Architecture for AI Agent Action Accountability

As organizations deploy AI agents to execute tasks in business processes, to our knowledge, no systematic academic framework exists for recording, verifying, and ensuring accountability for their actions. This study surveys industrial challenges and regulatory trends in AI agent action recording, formalizes a threat model (T1–T8), and constructs a seven-dimension evaluation framework (D1–D7). Through systematic survey of 16 academic papers and 5 industry projects and comparison matrix analysis, it identifies research gaps and demonstrates the non-substitutability of architectural differences against the two closest existing works (PROV-AGENT, InALign). The survey confirms that "a cryptographically verifiable, independent third-party recording standard for AI agent runtime behavior across platforms" remains an open research problem not addressed by the surveyed works. This paper proposes AEGIS (Agent Execution Governance and Integrity Standard), a three-layer reference architecture targeting these requirements, and evaluates it through a reference implementation across metrics M1–M6 and scenarios S1–S5.