Enterprise AI platform · 2025 –
Three-Plane Enterprise Agent Platform
Senior AI Engineer, Platform
Challenge
The organization needed an enterprise-grade foundation for multi-tenant AI agents: marketplace and identity on one plane, LLM/RAG/memory on another, tool routing and workflow execution on a third. Planes had to stay decoupled so isolation, releases, and onboarding did not break.
Approach
- Designed a three-plane architecture: control plane (marketplace, identity, agent catalog), data plane (LLM gateway, RAG service, memory), execution plane (API gateway, MCP/tool routing, Temporal workflow workers, agent runtimes).
- Standardized delivery on GKE, Helm, and GitOps across environments.
- Defined MCP and gateway routing patterns so tools and agents compose without bypassing tenant or auth boundaries.
Tech stack
TypeScript · Fastify · Temporal · Prisma · MCP · Kubernetes · Helm · GitOps · GCP (GKE) · Postgres · OIDC/JWT
Outcomes
- Clear plane boundaries for independent scaling and safer releases
- Multi-tenant enterprise agents deployable on Kubernetes with environment promotion via GitOps
- Foundation for RAG, evaluation, and customer workflow programs on a single platform
Verifiable patterns (redacted)
Plane boundary diagram (redacted)
Three-plane separation: control plane handles marketplace, identity, and agent catalog; data plane hosts LLM gateway, RAG, and memory services; execution plane routes MCP tools, API gateway traffic, and Temporal workers. Auth tokens never cross plane boundaries without explicit gateway validation.
GitOps promote manifest (redacted)
Helm values per environment with plane-scoped release tags. Promote pipeline blocks if execution-plane image bumps without matching data-plane contract version — prevents silent retrieval or tool-schema drift across tenants.
MCP gateway routing (redacted)
Tool dispatch table: agent runtime → execution gateway → MCP server, with tenant ID and execution token injected at each hop. Tools registered per org; cross-tenant tool invocation returns 403 at gateway, not at tool runtime.
Context
Internal platform serving multiple enterprise customer programs on shared infrastructure.