Output-Maximizing Long-Context Programming: Why Agentic Coding Workflows Scale Poorly and What to Do Instead

Current large language model development tooling is dominated by agentic, iterative workflows that emphasize small outputs, frequent context rehydration, and substantial invisible overhead. This paper argues that these workflows are largely an artifact of historical output constraints rather than an inherent requirement of intelligent systems. It introduces Output-Maximizing Long-Context Programming (OMLCP), a paradigm that exploits large output windows to generate coherent, large-scale artifacts in single or near-single constrained executions. Through formal token-economics analysis, sensitivity analysis, and field evidence from three case studies, the paper argues that agentic workflows exhibit severe compounding inefficiencies driven primarily by non-output overhead, especially repeated context rehydration. In contrast, OMLCP preserves global coherence and can reduce total token cost by one to two orders of magnitude for well-specified tasks. A full-stack application case study is presented in which 14,431 lines of code were generated at a total cost of $0.58, representing an estimated 26–52× efficiency advantage over comparable agentic approaches. The paper also includes a reproducibility protocol for independent validation and a capability-tier framework designed to reduce dependence on specific model versions.