Instruction Bleed Causes Cross-Module Interference in Agentic Systems.

A significant and often overlooked failure mode in prompt-composed agentic systems, termed "instruction bleed" or compositional behavioral leakage (CBL), has been formalized. This phenomenon occurs when modifications to one prompt module inadvertently alter the behavior of other modules, even without direct shared variables or explicit dependencies. The root cause is identified as the architectural non-isolation inherent in transformer self-attention mechanisms, which lack formal boundaries between concatenated modules within a context window. Researchers probed CBL using a deployed job-evaluation agent, conducting 144 trials with Claude Sonnet 4.6. A three-channel protocol was employed to perturb non-focal modules along volume, content, and form. The study found that only changes in the content channel produced a detectable, albeit subtle, paired effect. While no major recommendations were flipped, this sub-threshold regime of interference can accumulate across thousands of decisions made by a deployed agent, leading to significant cumulative impact. This work contributes an operational definition, a reusable protocol, and a falsifiable prediction set for CBL. It establishes cross-module interference measurement as a critical requirement for the robust evaluation of prompt-composed agentic systems, distinguishing it from other known agent failure modes like adversarial injection or cognitive degradation.