Data Facts: A Metadata Schema for Structured Data Exchange in the NANDini Multi-Agent Ecosystem

Jin Gao, Maria Gorskikh, Pradyumna Chari, Brittany Box, Mukul Kemla, Pratik Behera, Abhishek Mehta, Ramesh RaskararXiv 2026 · arXiv

TL;DR

This paper introduces Data Facts, a lightweight metadata schema and security pipeline that allows autonomous AI agents to discover, verify, and securely exchange structured datasets. By bridging the gap between agent communication protocols and data governance, it enables multi-agent ecosystems to make 100% accurate, data-driven decisions in real time without human intervention.

What Problem Does This Paper Address?

As autonomous multi-agent systems scale to billions of nodes, agents must independently generate, refine, and exchange data to make decisions. Current agent infrastructure protocols (like NANDA for registries, or Google A2A and Anthropic MCP for messaging) excel at identity and tool invocation but completely lack a machine-readable layer to describe, verify, or secure the datasets agents hold.

Conversely, existing enterprise data-sharing frameworks (such as Gaia-X or Ocean Protocol) rely on heavyweight, human-in-the-loop governance and negotiation cycles that are too slow for real-time agent interactions. Without an automated way to evaluate data freshness, integrity, and access permissions, agents are forced to rely on hard-coded endpoints or fall back on stale training priors, leading to inaccurate decisions and hallucinations.

Methods

The authors propose Data Facts, a core extension to the NANDini multi-agent framework. To keep agent registries lightweight, Data Facts introduces a single pointer—data_facts_url—to the existing Agent Facts registry record. This URL resolves on-demand to an external JSON metadata document specifying dataset identity, access tiers (public, semi_private, or private), endpoints, a time-to-live (TTL) window for freshness, and a SHA-256 integrity checksum.

The architecture is split into two coordinated layers:

  1. Agent Layer: Handles discovery and client-side validation of data freshness and integrity.
  2. Data-Owner Layer: Features a Data Layer Gateway service that acts as an adapter in front of heterogeneous databases (relational, vector, graph, etc.). For non-public data, it enforces a three-layer security pipeline: JWT authentication, capability-scoped gateway authorization, and an Agent-to-Agent (A2A) credential delegation protocol.
Figure 1: NANDini Data Facts architecture: agent layer (discovery, communication) and data-owner layer (gateway enforcement, capability delegation).

Datasets

The framework was evaluated using a newly designed benchmark consisting of 840 decision-making evaluations across seven binary decision scenarios. These scenarios were drawn from a PostgreSQL database containing a ground-truth distribution of items (labeled Gold, Silver, and Bronze).

The queries were specifically engineered to be counter-intuitive (e.g., asking for exact counts or parity, like the near_parity control scenario) to expose whether an LLM agent was guessing based on its training priors ("self-decide" mode) or querying live data ("data-informed" mode).

Key Insights and Ideas

  • The Latency Inversion Paradox: Intuitively, adding a database query step should increase latency. However, the paper reveals that for hard, counter-intuitive questions, data-informed agents actually achieved lower tail (p95) latency than self-deciding agents. Without data access, the LLM spends significant "deliberation cost" trying to reason through an impossible problem, generating verbose, slow completions. Constraining the agent with structured data access drastically cuts down output token length, saving time.
  • Registry Separation of Concerns: By using an external URL pointer rather than embedding dataset details directly into the registry, metadata can evolve independently without inflating the central registry payload or slowing down basic agent discovery.
  • Claim Validation is Mandatory: The security evaluation proved that signature verification alone is insufficient for autonomous systems. Attackers can present a validly signed token but alter the internal dataset_id claim. Gateways must explicitly validate claims against the requested capability to block sophisticated forgery.

Key Results

  • 100% Decision Accuracy: Data-informed agents achieved perfect accuracy (420/420 trials) across all agent topologies (1-agent direct gateway, 2-agent delegation, and 3-agent relay chain), compared to just 35.2% accuracy for agents relying on training priors alone.
  • Freshness & Integrity Enforced: Client-side TTL enforcement reduced stale-data decision errors from 37.6% to 8.8%. Recomputing SHA-256 checksums successfully detected 100% of data corruption events, even at a minimal 1% byte-level injection rate.
  • Bulletproof Security: The three-layer security pipeline blocked all 46 adversarial forgery attempts across 13 attack categories with zero data leakage.
Figure 2: Decision quality on data-driven questions with increasing difficulty.

Limitations and Future Work

  • Evaluation Scope: All decision-making evaluations relied on a single model (gpt-5-nano) and a binary classification format; the results may not immediately generalize to open-ended queries or other model families.
  • Token Reuse Vulnerability: In this v1 implementation, the stateless JWT pipeline does not bind tokens to a specific caller identity. This allows an intercepted token to be reused within its valid TTL window.
  • Future Extensions: Immediate future work includes implementing token revocation registries, audience binding, and short-lived credential exchanges to fix the identity-binding limitation. The authors also plan to introduce a catalog endpoint that lists multiple datasets per agent to minimize network round-trips.

Summary updated

Abstract

NANDini (Networked Agents Natural Distillation of Interconnected Nodal Intelligence) envisions an automated ecosystem where intelligent agents independently create, process, and exchange data to drive decisions at scale. Realizing this vision requires infrastructure beyond agent discovery and communication: agents must be able to advertise, evaluate, and verify the datasets they hold. Current protocols, including NANDA for federated registry and A2A and MCP for inter-agent messaging, address identity and communication but provide no mechanism for structured data exchange. Existing enterprise data-sharing frameworks, such as IDS-RAM, Gaia-X, and Ocean Protocol, assume human-in-the-loop governance that is incompatible with autonomous, real-time agent interactions. We introduce Data Facts, a core NANDini concept: a lightweight JSON metadata schema that bridges agent discovery and data access through a single pointer, `data_facts_url`, added to an existing Agent Facts registry record. The linked document encodes dataset identity, access tier, whether public, semi-private, or private, endpoint, a time-to-live for freshness validation, and a SHA-256 integrity checksum. For private and semi-private data, we implement a three-layer security pipeline: JWT authentication, capability-scoped gateway authorization, and an A2A credential delegation protocol. Across 840 decision-making evaluations, data-informed agents achieve 100% accuracy versus 35.2% without data access (p < 0.001); TTL enforcement reduces stale-data errors from 37.6% to 8.8%; checksum verification achieves 100% corruption detection at all injection rates; and the security pipeline blocks all 46 forgery attempts with zero data leakage.
BibTeX
@misc{gao2026data,
  author = {Jin Gao and Maria Gorskikh and Pradyumna Chari and Brittany Box and Mukul Kemla and Pratik Behera and Abhishek Mehta and Ramesh Raskar},
  title = {{Data Facts: A Metadata Schema for Structured Data Exchange in the NANDini Multi-Agent Ecosystem}},
  year = {2026},
  eprint = {2606.26211},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2606.26211},
}
Figure