The proliferation of Large Language Model (LLM) agents has been predicated on an assumption of "Abundance Connectivity" -- the idea that high-bandwidth, low-latency internet access is ubiquitous and continuous. Frameworks such as LangChain and AutoGPT operate on synchronous request-response cycles that fail catastrophically when network stability fluctuates. In the Global South, where intermittent connectivity and high latency are architectural constraints rather than edge cases, this creates an "Agentic Gap": the divergence between an agent's theoretical capability and its operational reliability. This paper introduces Contextual Engineering, a reference architecture that decouples agentic reasoning from immediate network availability. By implementing "Offline-First" state management and hybrid inference routing, we demonstrate that agentic systems can achieve high reliability in hostile infrastructure environments without sacrificing model intelligence.