An Event-Driven Software Architecture for Situated LLM Agents

Realizing the potential of Agentic AI requires agents to maintain a continuous, structural coupling with dynamic environments. Mainstream paradigms, however, often reduce this complex interaction to synchronous, stateless function calls, a simplification that proves inadequate for persistent, asynchronous domains. This thesis addresses a fundamental engineering challenge: establishing the architectural prerequisites that enable agents based on Large Language Models (LLMs) to effectively operate tools as persistent, observable, and asynchronous entities. To resolve this challenge, we introduce the Apprentice framework. Rooted in the Agents & Artifacts (A&A) theory, the framework introduces "Enhanced Tools" characterized by: active observability of state, asynchrony via event emission, and self-description through semantic manuals. Our analysis demonstrates a critical impedance mismatch when integrating these tools with standard reasoning loops like ReAct: synchronous implementations suffer from State Blindness and Execution Blocking, rendering them structurally insufficient for complex interaction. In response, we propose the S-ORA (Situate-Observe-Reason-Act) architecture, a decision cycle tailored to decouple context alignment from reasoning. Unlike traditional linear models, S-ORA allows the system to suspend specific Activities into a semantic "waiting" state without halting the global agent runtime, efficiently reconciling active state inspection with passive event handling. We validate this baseline through simulated scenarios: managing critical infrastructure and multi-agent coordination. The results demonstrate that S-ORA successfully handles temporal safety constraints and prevents infinite polling loops. This research confirms that the shift towards situated agency necessitates a dual evolution: the adoption of artifact-based tool design and the redesign of the agent’s temporal control flow and memory management systems.