In the domain of intelligent process automation, agents powered by Large Language Models (LLMs) are increasingly deployed for complex tasks because of their generative capabilities. However, existing architectures relying on finite context windows and fuzzy vector retrieval suffer from progressive loss of context and stochastic instability, which limits their reliability for long-horizon, regulated enterprise workflows. This thesis designs and evaluates a Goal-Oriented Knowledge Graph (GOKG) architecture to address this deficiency by externalizing agent state, memory, and process logic into a deterministic graph structure.

Following a Design Science Research (DSR) methodology, a hybrid neuro-symbolic prototype was developed and evaluated using comprehensive functional unit testing (N = 234 runs across 12 scenarios) and semi-structured expert interviews (N = 5). Results indicate that the graph-based state machine effectively enforced structural determinism, achieving a 100% success rate in maintaining process logic and state persistence across the tested scenarios. Furthermore, the system demonstrated over 90% accuracy in complex knowledge extraction tasks, significantly outperforming purely prompt-based baselines. These findings suggest that decoupling the reasoning engine (LLM) from the state management layer (Graph) iscritical for auditability, implying that future enterprise AI systems should adopt “Auditable” architectures to ensure verifiable determinism.

Findings are limited by increased system latency and modeling complexity, indicating that future research should prioritize runtime optimization and low-code abstraction tools.