AgentNexus: A Service-Boundary-Aware Coordination Architecture for Heterogeneous LLM Code Agents

When a fleet of LLM code agents works on a large software system, the hard problem is not generating code but keeping distributed, independently evolving artifacts consistent across the teams that own them. We argue that the right unit of coordination is neither the agent nor the service but the ownership boundary: a region of independent consistency responsibility that owns a set of versioned documents. We make this primitive operational. An operational criterion — the subscription litmus test — decides where a boundary is drawn: two units are distinct boundaries exactly when a contract flows from one to the other. We separate ownership, which is singular and anchors consistency, from attestation — a self-attested role (a Principal) recorded on each write like a Git author — so that many actors can collaborate on one boundary without fracturing its ownership. We further show that propagating knowledge across boundaries is a channel problem, not a payload problem: document content flows out-of-band over HTTP (a data plane, at zero token cost to the model), while coordination signals flow over the Model Context Protocol (a control plane). We report a deployment on a commercial multi-service product with eleven boundaries, in which a read-only global observer surfaced a genuine cross-boundary configuration inconsistency. Supersedes v3 (doi.org/10.5281/zenodo.20603176). Source code: https://github.com/dugubuyan/agent-nexus