This paper introduces a state description for mobile agent systems on ad hoc networks, allowing agents to reason about how they communicate over the agent system?s underlying network. The ability to make intelligent decisions about communications allows agents to achieve their goals more efficiently, and increase the security and survivability of an agent system. One application of this model includes a formal representation of information assurance for agent messaging on dynamic networks with possibly redundant routes. We use this framework to study the "compromised host" problem — specifically, how agents should react after an intruder has been detected. This is done by first quantifying the effect of the compromised host on the integrity of messages among agents; and then selecting a routing policy with respect to message integrity and the compromised host (e.g., rerouting to avoid the host or continuing to communicate through it). We provide examples of how this technique can be used to improve information assurance for agent-to-agent communications and information distribution in sensor networks. Empirical validation in the SecureWireless Agent Testbed (SWAT) shows network awareness improves the survivability and efficiency of mobile agent systems.