Discussion draft / mesh-of-meshes study
A proposed control plane for the seams between service mesh, data mesh, and the emerging agent mesh.
Thesis
A mesh of meshes should not become a fourth mesh. It should be a thin governance-and-identity fabric that makes service, data, and agent meshes interoperable through governed interactions.
Every human, service, data, or agent call carries identity, delegation, purpose, a policy decision, and a correlated audit event. Governance is coordinated centrally and enforced locally.
Five pillars
The design constraint is practical: centralize identity, policy, and telemetry; keep enforcement close to the workload, data product, or tool broker.
Every human, service, data product, and agent is a verifiable principal. Purpose and on-behalf-of delegation travel across the chain so an agent cannot act with more authority than the person or workflow that invoked it.
One purpose-aware authorization model can be distributed to service sidecars, data-product gateways, and agent tool brokers. The fabric coordinates decisions without becoming a runtime chokepoint.
APIs, events, data products, and agent tools need shared contracts and vocabulary. Student should mean the same thing to a service, projection, and governed AI tool.
A correlation, causation, and purpose trail follows the interaction so the institution can answer who did what, on whose authority, against which data, and why.
Agents are probabilistic and prompt-injectable, so the fabric scopes tools, limits blast radius, requires human approval for high-impact actions, and structurally separates private data from exfiltration paths.
Credibility map
WGU lighthouse
The most useful MVP is not a big-bang platform. It is one governed path: an AI tutor or mentor assistant reaching the Unified Student Object through delegated identity, purpose policy, lineage, and human approval.
Tutor agent to Unified Student Object with identity, purpose, policy, lineage, and human approval.
Self-serve identity issuance, policy SDKs, reusable audit hooks, and two or three more onboarded seams.
Domain-owned policies inside global guardrails, with shared contract and semantic registries.
Agent registry, blast-radius controls, injection monitoring, drift telemetry, and lifecycle governance.
Receipts
A concrete AI trust layer pattern: agents reach systems through scoped, audited tools rather than raw vendor credentials.
Domain-owned operational truth exposed through event streams, projections, contracts, and anti-corruption adapters.
A governed software-factory pattern: quality gates, provenance, human review, and auditable delivery for AI-assisted engineering.
Target shape
Thin federated overlay: the recommended core because it reuses existing mesh investment and keeps enforcement local.
Portable identity spine: purpose-carrying identity tokens become the backbone for policy and audit decisions.
Governance broker for high-risk agent traffic: the Sentinel pattern, kept off latency-critical hot paths.
Policy at the data: student-data products enforce identity and purpose at access time.
Adoption
Earn adoption seam by seam instead of mandating a platform rewrite.
Keep brokers off latency-critical paths and make local enforcement the norm.
Treat prompt injection as a condition to contain and monitor, not a solved problem.
Curate contracts and semantics continuously because they drift when ownership gets vague.
BMOZI Technical