Enabling AI Agent Negotiation in Data Spaces

Abstract

Data Spaces are federated ecosystems for cross-institutional data sharing that preserve each participant's sovereignty over how their data is accessed and used. Their underlying software components, however, are built for deterministic machine-to-machine automation and offer no capacity to reason or adapt when interactions fall outside anticipated scenarios, a limitation that becomes especially apparent when negotiating access under changing or time-sensitive conditions. This dissertation addresses that gap. Following an Engineering Research methodology and grounded in health-sector Data Space work at INESC TEC, it designs the Data Space Agent Negotiation Protocol: a bilateral, alternating-offer protocol in which agents exchange complete ODRL policy offers and reason privately over machine-readable governance profiles. By operating through the Data Space infrastructure rather than around it, the protocol ensures that the terms reasoned over and those ultimately enforced are the same objects. To support this, a deterministic policy engine retains final authority over agreements, while a large language model is confined to proposing within bounds the engine re-validates. Ultimately, the protocol, an autonomous agent built to enact it, and a human-facing dashboard are combined into a working proof-of-concept. The system is evaluated through a capability rule-book benchmark over 270 autonomous simulations, complemented by an ablation with the language model disabled. Every capability rule held on every run, demonstrating consistently correct and repeatable behaviour across the full range of participant situations: negotiating to agreement, fetching directly, refusing on a red line, and abandoning unsatisfiable tasks. The language model reasoned autonomously yet never breached its deterministic bounds, and every outcome remained a bounded, auditable function of the parties' declared policies. This work delivers an engineered foundation for the next generation of agent-driven, sovereignty-preserving data sharing, and a precise account of the boundaries that future work can build beyond.