Architectures for Building Agentic AI

Chapter 3: Architectures for Building Agentic AI Slawomir Nowaczyk[0000−0002−7796−5201] Abstract This chapter argues that the reliability of agentic and generative AI is chiefly an architectural property. We define agentic systems as goal-directed, tool-using deci- sion makers operating in closed loops, and show how reliability emerges from princi- pled componentisation (goal manager, planner, tool-router, executor, memory, verifiers, safety monitor, telemetry), disciplined interfaces (schema-constrained, validated, least- privilege tool calls), and explicit control and assurance loops. Building on classical foundations, we propose a practical taxonomy—tool-using agents, memory-augmented agents, planning and self-improvement agents, multi-agent systems, and embodied or web agents—and analyse how each pattern reshapes the reliability envelope and fail- ure modes. We distil design guidance on typed schemas, idempotency, permissioning, transactional semantics, memory provenance and hygiene, runtime governance (bud- gets, termination conditions), and simulate-before-actuate safeguards. 1 Introduction: purpose, scope, and architecture reliability This chapter surveys architectural choices for building agentic AI systems and analyses how those choices shape reliability. Our central claim is straightforward: reliability is, first and foremost, an architectural property. It emerges from how we decompose a system into components, how we specify and enforce interfaces between them, and how we embed control and assurance loops around the parts that reason, remember, and act. Individual models matter, but without the right architectural scaffolding, even state-of-the-art models will behave inconsistently, be impossible to audit, and prove fragile in the face of novelty. This is a preprint of a chapter accepted for publication in Generative and Agentic AI Reliability: Architectures, Challenges, and Trust for Autonomous Systems, published by Springer Nature. Agentic AI in this book denotes systems that pursue goals over time by deciding what to do next, selecting and using tools, consulting and updating memory, and inter- Slawomir Nowaczyk Center for Applied Intelligent Systems Research, Halmstad University, Sweden e-mail: sla- womir.nowaczyk@hh.se 1 arXiv:2512.09458v1 [cs.AI] 10 Dec 2025 2 Slawomir Nowaczyk acting with their environment under constraints. An agent is not merely a predictor; it is a decision-maker in a closed loop. It observes, plans (or at least chooses), acts, and learns, typically under uncertainty and partial observability. Generative AI refers to models that synthesise content—text, code, images, plans, or intermediate representa- tions—often serving as the reasoning substrate inside the agent, or providing artefacts (queries, programs, simulations, explanations) that other components execute or verify. In modern systems, generative models supply the policy (how to reason and propose actions), while the agentic architecture supplies the machinery (how proposals are validated, enacted, bounded, and recorded). Understanding the relation of Agentic GenAI with classic autonomous agents is crucially important to avoid reinventing the wheel: many key concepts have been studied for a long time and are relatively well-understood today; however, the nature of GenAI also brings up challenges that are completely novel and require rethinking of what was believed to be known. Traditional reactive, deliberative, or BDI (belief- desire-intention) architectures offer theoretically-founded and crisp notions of concepts such as beliefs, goals, plans, and intentions, with clear control loops and explicit world models. Modern agentic systems often replace hand-engineered reasoning with neural- network-based foundation models. These models, trained on huge amounts of diverse data, vastly increase the flexibility and breadth of competence, but also introduce uncertainty in reasoning steps and tool usage. In this chapter, we retain the useful discipline of the classic view—explicit state, goals, plans, commitment strategies, and monitoring—while acknowledging that parts of the pipeline (e.g., plan generation or hypothesis formation) may be implemented by generative models. That reconciliation is precisely where architecture earns its keep. This book is not intended as yet another broad introduction to Agentic GenAI; instead, we put these recent developments in the specific context of reliability. By reliability, we mean the consistent achievement of intended outcomes under stated conditions, within acceptable bounds of safety, security, data protection, and resource usage, and with evidence that failure modes are known, contained, and recoverable. For agentic AI, this encompasses much more than just model accuracy. It includes correct tool invocation, bounded action sequences, resistance to manipulation, predictable latency and cost, graceful degradation, auditability, and human-override paths. Architectures make these

This content is AI-processed based on open access ArXiv data.