Guard System Overview

These components work in sequence to sanitize, assess, and shape communications, as depicted inthe interaction pipeline.

The guard system role in agent communication is illustrated in Figure 2, which shows the sequential flow: input message passes through the input filter and pre-inference guard model check, reaches the base model (potentially with guard prompt), and the output undergoes post-inference angel check and output filter before delivery. This pipeline ensures multi-stage protection, reducing the risk of successful attacks. In practice, the guard layer adds minimal latency (e.g., <100ms for filters, <500ms for angel model on efficient hardware) while significantly enhancing security.