AATF – An open spec for recording why AI agents make decisions (github.com)

The AI/ML community was introduced to the AATF (AI Agent Traceability Framework), an open specification designed to document the reasoning behind AI agent decisions, including considered alternatives and confidence levels. Unlike traditional logging libraries, AATF focuses on accountability by capturing detailed decision-making processes of AI agents. For instance, when a user requests a flight to Shanghai, AATF records every step—from identifying the intent to rejecting alternatives—providing a structured, tamper-evident audit trail. This development holds significant implications for transparency in AI systems, as it allows developers to ensure that their agents make well-reasoned decisions while adhering to compliance standards like GDPR and the EU AI Act. With features such as numeric confidence scoring, alternative consideration tracking, and PII redaction, AATF empowers developers, compliance officers, and researchers by offering insights into decision patterns and enhancing the integrity of AI applications. AATF is a community-driven initiative, inviting feedback and contributions to refine its specifications further, indicating a progressive step toward responsible AI usage.