What building Shippy taught us about building agents

Hugging Face · 3 months ago · 33 ● 2 sources

Shippy, an AI agent for maritime domain awareness, was built with a three-part architecture—system prompt ("soul"), task-specific instructions ("skills"), and configuration settings—designed to prioritize reliability over raw model capability in high-stakes operational decisions. The system includes a purpose-built CLI layer that standardizes API calls to prevent subtle bugs, isolated user sandboxes running on Kubernetes for data privacy, and a custom evaluation framework that scores the entire agent against weighted criteria on live data rather than static benchmarks. This approach enables Shippy to serve 70+ countries while maintaining data isolation between users and prevents deployment of versions that regress on guardrails or accuracy thresholds.