Sovereign Transaction Graphs – Adaptive Agents Without Context Leakage

Session track: AI Engineering

Session time: 01:30 pm - 02:00 pm UTC November 12

Session description:

Autonomous agents are increasingly expected to act on behalf of people and organizations in transactional settings, such as price negotiation, service coordination, procurement, or data sharing. To do so well, they require information about the participants’ goals, needs, constraints, preferences, and risk tolerance. However, this information is often sensitive. If it is simply added to an agent prompt, the very data required for context-adaptive interaction may become a source of leakage, misuse, or mandate violations. In this session, Mandy Goram will present an MVP of a broader framework and platform for context-adaptive agent interaction. The session will focus on the use of Neo4j as a runtime context layer for participant-owned context graphs. Each participant will maintain a private graph containing needs, goals, constraints, preferences, mandates, risk profiles, and disclosure rules. A local context broker will use this graph to derive a safe interaction envelope: the limited set of facts, actions, proposals, and disclosures an agent may use in the next transaction step. You will learn how to model such participant-owned context in Neo4j, how Cypher queries can support mandate and disclosure checks, how a Go-based control loop can adapt agent behavior at runtime, and how decision traces can make autonomous actions auditable. A marketplace-style price negotiation will serve as the concrete use case, while the underlying pattern will be applicable to broader multi-party transactions. The key takeaway will be that context-adaptive agents should not be built by exposing more private context to prompts, but by designing systems that use context while protecting the participants it describes.