D-MARN-RT: Real-Time Decentralized Negotiation for Agentic Commerce
This research examines D-MARN-RT, an extension of the D-MARN framework for autonomous value discovery and negotiation in agentic commerce. The work investigates how buyer and merchant agents can discover compatible offers, enforce hard constraints, negotiate transaction terms, and produce verified commitments before checkout and payment.
The question
Emerging agentic-commerce protocols address important parts of autonomous transactions, but they do not fully specify how agents should discover and negotiate value before checkout. UCP can standardize commerce capabilities and transaction operations, A2A can support agent communication, MCP can provide access to tools and context, and AP2 can bind authorization to payment. A gap remains between expressing a user's intent and reaching a transaction-ready commitment with a merchant.
The central research question is:
How can a decentralized, protocol-aware multi-agent architecture transform structured buyer intent into a constraint-safe, negotiated transaction commitment while remaining scalable, auditable, and compatible with emerging commerce protocols?
Motivation
Conventional e-commerce assumes that users search, compare, and select from largely static offers. Agentic commerce changes this interaction model. A software agent may need to interpret a goal, distinguish hard requirements from preferences, discover heterogeneous sellers, compare multiple attributes, negotiate terms, and verify that the final offer remains within authorization boundaries.
The challenge is not simply ranking similar products. A highly relevant offer may still violate a budget, privacy rule, delivery requirement, or organizational policy. For this reason, the research separates deterministic constraint enforcement from probabilistic semantic reasoning.
Scope
The current work focuses on decentralized candidate discovery, structured intent representation, hard-constraint gating, multi-attribute evaluation, bounded negotiation, commitment verification, and transaction handoff. D-MARN-RT extends this direction toward real-time bargaining and protocol-aware implementation.
Open gaps include negotiation scalability, transport latency, strategic seller behavior, incentive compatibility, privacy-preserving intent disclosure, and validation using realistic merchant systems rather than only synthetic marketplaces.
The approach
The completed D-MARN foundation defines Decentralized Multi-Agent Resource Negotiation as an upstream value-discovery and negotiation layer for agentic commerce. The framework separates semantic reasoning from deterministic policy enforcement.
Buyer intent is represented using a UCP-Encoded Intent Tensor, or UEIT. The representation separates semantic objectives, preference weights, hard constraints, context, authorization boundaries, and trust or risk requirements. Candidate sellers are first discovered broadly and then filtered using deterministic feasibility checks. Only feasible candidates proceed to scoring and negotiation.
The completed D-MARN workflow includes:
- intent compilation into a structured UEIT;
- capability and offer discovery;
- deterministic hard-constraint gating;
- personalized multi-attribute utility estimation;
- bounded local negotiation with candidate sellers;
- final commitment verification; and
- handoff of the verified commitment to UCP checkout and AP2 authorization.
The architecture was evaluated using a synthetic marketplace containing heterogeneous seller attributes and buyer preferences across different marketplace sizes.
D-MARN-RT work in progress
D-MARN-RT represents the next stage of the research. The work explores a more explicitly real-time negotiation architecture in which buyer and seller agents can exchange proposals and counteroffers with lower coordination overhead.
Planned work includes evaluating streaming communication approaches such as gRPC and Protocol Buffers while maintaining compatibility with higher-level agent and commerce protocols. The research also explores game-theoretic bargaining mechanisms, including Nash-bargaining-style objectives, as potential negotiation policies. These mechanisms are proposed extensions and are not yet established findings.
Evaluation plan
Planned evaluation criteria include:
- realized buyer utility;
- negotiated price and economic surplus;
- hard-constraint violation rate;
- no-match rate;
- messages per transaction;
- end-to-end negotiation latency;
- throughput as seller population increases;
- sensitivity to candidate-set size;
- ablations for negotiation, personalization, and constraint gating; and
- statistical comparisons across multiple random seeds.
The planned reproducibility package includes the simulator, experiment configurations, random seeds, raw outputs, and figure-generation scripts.
Limitations
The current evidence is primarily synthetic. It does not yet establish production latency, merchant revenue improvement, incentive compatibility, or real-world commercial effectiveness. Real UCP-compatible merchant environments, adversarial agents, privacy-preserving negotiation, formal safety verification, and live transaction pilots remain future validation steps.
What it explores
The current experimental evidence applies to the original D-MARN framework rather than the full D-MARN-RT extension.
In a synthetic marketplace with 100 seller agents evaluated across five seeds, D-MARN achieved a mean realized utility of 0.734, compared with 0.714 for a scalar-utility baseline and 0.737 for a centralized exhaustive oracle. The experiment therefore placed D-MARN close to the centralized upper-bound result while requiring substantially less coordination.
D-MARN used an average of 32.2 messages per buyer intent compared with 100 messages for the centralized oracle, representing approximately a 67.8% reduction in coordination traffic in that experiment.
The architecture also produced zero hard-constraint violations in the reported evaluation because feasibility was checked before selection and verified again before commitment. Simpler semantic and scalar baselines produced hard-constraint violations in a substantial fraction of matched outcomes.
What D-MARN-RT explores
D-MARN-RT investigates whether these architectural principles can be extended into a lower-latency, more explicitly real-time bargaining environment.
No current result establishes that gRPC, bidirectional streaming, Nash bargaining, or other proposed D-MARN-RT extensions improve production transaction performance. These remain research questions.
The next evaluation is intended to examine whether real-time negotiation can preserve constraint safety and near-optimal value while improving communication efficiency, scalability, and transaction responsiveness. Particular attention is given to the trade-off among negotiation quality, message cost, latency, decentralization, and merchant autonomy.