What building Shippy taught us about building agents

An AI agent built for maritime domain awareness was designed with three core components: a system prompt that defines its behavior, specific skills for handling particular requests, and configuration settings that determine which language model and framework it uses. The agent's reliability depended not on the language model itself, but on building a system that could be trusted to stay within its limits and provide correct answers, since wrong information in a high-stakes operational domain could have serious real-world consequences. The team implemented several key architectural decisions to ensure predictability, including a purpose-built command-line interface that simplifies complex API interactions and reduces the agent's ability to make subtle errors, as well as isolated user sessions that keep each person's data and conversation history separate. These design choices, where each system layer can be tested independently and behavioral boundaries are explicitly defined rather than left implicit, offer lessons for building other AI agents in high-stakes domains.