Large language models (LLMs) exhibit degraded constraint compliance under prompt compression, with violations peaking at medium-length instructions. While this phenomenon has been documented in base models through the Compression-Decay Comprehension Test (CDCT), it remains unknown whether deployed agentic systems:LLMs equipped with tool access, memory,and multi-step planning:exhibit similar failure modes. We introduce CDCT-A, an adaptation that measures constraint compliance in agents across compressed tool documentation, finding that agents display an asymmetric U-curve with a 28.6 percentage point drop at medium compression (c=0.5: 63.3% vs c=1.0: 91.9%). More critically, we identify a dominant failure mode:vocabulary hallucination:where agents correctly infer user intent but substitute plausiblebut-incorrect API syntax, leading to 29.6% successful violations at medium compression: executions that complete tasks while violating specifications. Analysis of 1,350 evaluations across 9 frontier models and 10 tasks reveals that (1) agent scaffolding mitigates but does not eliminate compression failures, (2) vocabulary errors are deterministic and universal across all tested models under temperature=0, (3) model performance varies substantially (26.7pp range) in ways consistent with architectural differences in schema validation, and (4) standard task success metrics are blind to 20.6% of constraint violations. These findings demonstrate that compressioninduced failures generalize from base models to deployed agents, but manifest through different mechanisms requiring distinct evaluation and mitigation strategies.