Context Rot: Why MCP is Breaking in Production and How the Skills Architecture Fixes It

As large language model (LLM) agents are deployed at production scale, a class of system-level failure has emerged that we formally designate context rot-the progressive degradation of reasoning quality, task coherence, and decision fidelity arising from uncontrolled accumulation, misalignment, and entropic decay of context across extended agentic sessions. Although the Model Context Protocol (MCP) provides a standardized interface for tool invocation and external state retrieval, practitioner-reported evidence and our controlled simulation benchmarks reveal systematic breakdown patterns under conditions of scale, multi-agent coordination, and long-horizon task execution. This paper presents: (a) a formal taxonomy of context rot failure modes, (b) an information-theoretic framing of context as a bandwidth-limited computational resource, and (c) a comparative evaluation of MCP-based versus Skills Architecture systems across five benchmark domains. In controlled experiments using calibrated simulation over the Agentic Context Benchmark (ACB-2026), Skills Architecture reduces normalized context entropy by 42% (95% CI: [37%, 46%]) relative to baseline MCP, improves task completion rates by 26 percentage points on average (95% CI: [20pp, 31pp]) across five domains, and reduces error recurrence rates from 0.34 to 0.09. Results are consistent across three model families, supporting the position that context rot is architectural in character to a degree that survives cross-model comparison, though we do not claim to have isolated architecture as the sole cause. The full benchmark suite, reference implementations, and analysis code are publicly available for independent replication.