Problem

The core problem addressed is the lack of reliability in agentic AI systems, where even advanced generative models can behave inconsistently, be difficult to audit, and prove fragile when encountering novel situations. The paper argues that this unreliability stems not from the models themselves, but fundamentally from inadequate architectural design, including how systems are decomposed, how components interact through interfaces, and the absence of robust control and assurance loops.

Method

The paper surveys and analyzes architectural choices for building reliable agentic AI systems. It emphasizes three foundational design principles: effective componentization to isolate responsibilities and contain faults; well-defined interfaces and contracts to constrain open-ended model behavior; and robust control and assurance loops to govern generative components. It reconciles classic agent architectures (reactive, deliberative, hybrid, and BDI) with modern generative AI, detailing how concepts like beliefs, desires, and intentions can be mapped onto contemporary systems. The chapter then categorizes modern designs into five families—tool-using, memory-augmented, planning and self-improvement, multi-agent, and embodied/web agents—explaining the reliability implications and common failure modes for each. Key engineering choices like typed schemas, idempotent tools, capability-based permissioning, transactional semantics, budgets, and deterministic observability are presented as crucial for designing in reliability.

Results

As a survey and conceptual analysis, the paper does not present empirical results in the traditional sense. Instead, it provides a comprehensive framework demonstrating how specific architectural patterns and engineering practices contribute to agentic AI reliability. Through a running example of a tool-using diagnosis agent in a safety-critical environment, it illustrates how componentization, structured interfaces, and control loops enable containment, least authority, validation before actuation, assured fallbacks, graceful degradation, observability, auditability, memory hygiene, and cost governance. The discussion of various architectural patterns (e.g., MRKL, ReAct, ReWOO, ToT, GoT, PAL, Reflexion) shows how each can enhance or challenge reliability, offering concrete strategies to mitigate risks like hallucinated tools, infinite loops, and state explosion, thereby making powerful generative models behave as dependable, bounded, and governable systems.

Implications

The paper's primary implication is that reliability must be a design-time consideration for agentic AI, driven by architectural choices rather than merely model selection. It provides a principled taxonomy and practical templates for engineers and researchers to build more dependable autonomous systems. By highlighting the critical role of component isolation, structured interfaces, and explicit control loops, the work sets a foundation for future development, ensuring that advanced generative capabilities are channeled into robust, auditable, and governable problem-solving. It underscores the need for treating aspects like memory, tool interaction, and multi-agent coordination as first-class architectural subsystems with defined policies and mechanisms for assurance.