Every developer building with AI agents eventually hits a question they can't answer: why did the agent do that? When agents delegate across multi-hop chains to negotiate, route, or approve, the reasoning behind each decision is never stored as data, so it can't be queried, audited, or trusted.
In this session, I'll show you how to close that gap with a decision-trace layer built on a Neo4j property graph. Every agent action becomes a first-class node, with inputs, reasoning steps, tool calls, and outputs connected by typed edges like DECIDED_BECAUSE and PRECEDED_BY that span agent boundaries.
You'll see the graph data model, the Cypher that rebuilds a full multi-hop causal chain in a single traversal, and why a property graph answers "why" where flat logs and OpenTelemetry can't. I'll run a live three-agent demo in which an approval agent queries its own past decisions as precedent before acting: graph-native explainability you can watch happen.
You'll leave able to model agent decisions as a graph, write the Cypher to audit them, and turn decision history into a behavioral trust score. The reference implementation is open source (Apache 2.0) and built on MCP and A2A, so you can instrument your own agents with a single decorator.
Speaker
Alexis Ngoga
AI/ML | Agentic AI Systems| Solution Architect
Ngoga Alexis is an AI Engineer and Solutions Architect with a Master’s degree in Engineering Artificial Intelligence from Carnegie Mellon University. He currently works as a Data Engineer at RwandAir, where he integrates modern data and AI workflows with legacy enterprise systems. His experience spans AI agents, MLOps/LLMOps, and protocol-based integrations, with a focus on using open standards such as MCP, A2A, and WSDL to enable safe, incremental adoption of AI in production environments.
