Three-Ring Architecture offers governance for agentic AI to prevent 95% failure rate

A new paper from researchers Sergio Alvarez-Telena and Marta Diez-Fernandez tackles a structural crisis in enterprise AI: companies are deploying agentic AI without the governance infrastructure to control it. The Three-Ring Architecture proposes a formal solution to prevent the repeat of the first wave's 95% project failure rate, which was caused by decentralized intelligence without a federation layer. Ring 1 is the existing production architecture. Ring 2 is the critical innovation: a strategies-based agentic AI layer that functions exactly like an operating system for the enterprise—abstracting resources, coordinating processes, enforcing permissions, and providing a stable platform for compounding intelligence. Ring 3 contains LLM-based frontier intelligence, which is inherently non-deterministic and riskier.

The paper makes a central distinction between the risk profiles of Ring 2 and Ring 3. Strategies-based agents (Ring 2) operate deterministically: their actions are traceable, permissions enforceable, and deviations recoverable. LLM-based agents (Ring 3) introduce non-deterministic actors whose deviations can propagate through complex systems without retrospective traceability. The authors argue that Ring 2 is not optional—it is a necessary condition for control and compliance. Importantly, every improvement in LLM capability increases the governance requirement, as more capable non-deterministic actors produce larger consequences when they deviate. The architecture has been validated across a decade of deployment in financial services, government, procurement, and compliance sectors.