A FULL-STACK PRODUCTION-PLATFORM REFERENCE ARCHITECTURE FOR GOVERNED AGENTIC AND AUTOMATED-ACTION ENTERPRISE SYSTEMS

Abstract

A large fraction of enterprise AI systems that work in a pilot never operate reliably in production, because a running system must do far more than score a model: it must ingest heterogeneous data, turn it into analytics-driven decisions, execute actions, observe its own behavior, keep humans in the loop at the right moments, and improve from realized outcomes - continuously, under a service-level objective. This paper specifies a full-stack production-platform reference architecture for systems that plan, decide, and act, whether the actor is a conversational or LLM agent, a robotic or enterprise-process automation, an industrial or IoT control loop, an operational decision-support service, or an integrated analytics platform. Every layer is a first-class, contract-bearing component: a data foundation and shared context, an orchestration and workflow runtime with saga compensation, a residency-filtered model gateway and router, an analytics layer progressing descriptive-to-prescriptive, an action-and-tool-execution layer with transactional effects, and a human-in-the-loop review surface - wrapped by two cross-cutting spines, governance and an analytics-and-observability spine, and served by a control plane under an operational-readiness discipline. The observability spine is specified in depth: capture-at-commit execution records, a composite and agent-native validity signal, trajectory-level risk and credit assignment, drift-under-action, decision-trace completeness and freshness as operational measures, escalation under a capacity model, and a closed learning loop that retrains from realized outcomes and is defended as an attack surface. Seven normative invariants (R1–R7) span the stack, and three cumulative conformance profiles (Observable, Governed, Adaptive) let heterogeneous platforms adopt the specification incrementally and interoperate. The architecture is complementary to a decision-production and conformance framework such as Governed Decision-Intelligence (GDI): that framework certifies how each decision is produced; this one specifies how the system that acts on it is built, observed, and operated.