VRE Agentic Networks

[anormang1992]

Summary

Enable a network of VRE-bound agents that can query each other's epistemic graphs, sharing grounded knowledge across domain boundaries while preserving enforcement guarantees.

Problem Statement

The biggest bottleneck with VRE at the moment is epistemic graph curation. A single monolithic graph that covers every domain the agent might encounter doesn't scale — it requires exhaustive upfront authoring and becomes unwieldy as domains multiply.

A network model inverts this: each agent maintains a domain-specific graph (filesystem operations, network protocols, database management, etc.) and can query peer agents for grounded traces outside its own domain. This makes each graph smaller, more focused, and independently maintainable — effectively distributing the curation bottleneck across domain experts.

Proposed Solution

Network Topology

• Each agent in the network runs its own VRE instance backed by its own Neo4j graph
• Agents register with a discovery mechanism (service registry, config file, or peer-to-peer announce)
• When an agent encounters an ExistenceGap for a concept outside its domain, it can query the network for a peer that has that concept grounded

Cross-Graph Queries

• A querying agent sends a structured epistemic query (concept + required depth) to a peer
• The peer runs VRE.check() against its own graph and returns the GroundingResult (including the full epistemic trace)
• The querying agent incorporates the peer's trace into its own grounding decision — but does not persist the peer's knowledge into its own graph

Trust Boundaries

• Grounding traces from peers carry provenance metadata indicating the source agent and graph
• The consuming agent can apply trust policies: e.g. "accept grounding from fs-agent for filesystem concepts but not for network concepts"
• Trust is directional and scoped — agent A trusting agent B for domain X does not imply trust for domain Y

Protocol

• Request/response over a lightweight protocol (gRPC, HTTP/JSON, or message queue)
• Queries are stateless — no persistent connection between agents
• Responses include the full EpistemicResponse so the consumer can inspect the trace, not just the boolean result

VRE Design Alignment

• Agent–VRE contract preserved: Each agent still submits structured queries to its own VRE. Network queries are a fallback when local grounding fails — the contract interface doesn't change.
• Epistemic honesty across boundaries: A peer's grounded trace is not blindly trusted. The consuming agent knows the knowledge came from an external source (provenance) and can apply trust policies. The agent never appears more knowledgeable than it can justify — it can justify via its peer's trace, which is inspectable.
• No graph mutation: Cross-graph queries are read-only. The querying agent's graph is never modified by network responses. This preserves the integrity of each domain graph.
• Layered safety: Network trust policies are a new safety layer (analogous to Section 6's mechanical safety), sitting between epistemic safety (VRE) and human safety (approval layer).

Acceptance Criteria

• Agent discovery mechanism (at least config-file based for MVP)
• Cross-graph query protocol: agent can send epistemic query to a peer and receive GroundingResult
• Trust policy model: scoped, directional trust between agents
• Provenance metadata on network-sourced traces indicating source agent
• Consuming agent can incorporate peer traces into its own grounding decision
• No graph mutation from network queries
• Integration tests with two VRE instances exchanging grounded traces

Open Questions

• What is the MVP discovery mechanism — static config, DNS-SD, or a central registry?
• Should cross-graph results be cached locally (with TTL) or always queried live?
• How does network latency affect the grounding contract — should there be a timeout that fails open or fails closed?
• Can an agent delegate an entire grounding query to a peer, or only individual concept lookups?
• How does this interact with the confidence metric (Implement Per-Primitive Confidence-Based Epistemic Metric #2) — should peer-sourced knowledge carry a trust-discounted confidence score?

Dependencies

• Ensure Provenance Is a First-Class Attribute on Primitives and Relata #6 (Provenance as first-class attribute) — required to tag network-sourced knowledge with source agent
• VRE Informed Agentic Auto Learning #4 (Auto Learning) — network queries could feed the auto-learning pipeline when peers surface concepts the local agent lacks