AI Agent Audit Trails: Prove Why Your Agent Decided, Not Just What

Описание и применение

Создана система аудита решений ИИ, которая автоматически фиксирует реальные цепочки рассуждений и источники данных, используя Strands Agents и Neo4j. Это позволяет воспроизводить решения и находить все решения, затронувшие ошибочный источник данных.

Задача

AI agents often lack a reliable record of why they made a decision and which data sources influenced it, making it hard to audit and correct errors.

Результат

Implemented a zero‑change audit trail that records real reasoning chains and supports reverse audits in a single graph traversal, improving transparency and error correction.

Рабочий процесс

• 1. Register a HookProvider (DecisionTraceRecorder) on the Strands Agent. 2. Agent starts invocation; recorder opens a new trace. 3. For each tool call, AfterToolCallEvent is emitted; recorder records tool name, input, and source. 4. On invocation end, recorder closes trace with outcome. 5. Recorder writes trace to Neo4j via neo4j-agent-memory SDK. 6. Reverse audit: query Neo4j to find all traces that touched a given source.

Доказательства

Decision traces captured automatically with zero changes to your tools, and a reasoning graph, stored with Neo4j's official agent-memory SDK, where the reverse audit is a single traversal.

Источник

Your logs won't save you either. Logs record that things happened. An audit trail for an AI agent has to answer two harder questions…

Источник