Ai2's Shippy agent reveals blueprint for reliable AI in high-stakes ocean monitoring
A maritime AI agent that can't afford to be wrong: how Ai2 built Shippy with Claude Opus 4.6
Ai2's Skylight team built Shippy, an AI agent for maritime domain awareness where incorrect answers could waste patrol resources or endanger personnel. The agent's architecture separates three components: soul (a system prompt that defines behavioral boundaries), skills (markdown files following the same spec as coding tools like Claude Code), and config (harness, LLM choice, runtime secrets). Shippy currently uses Claude Opus 4.6 and the open-source OpenClaw framework.
Instead of letting the LLM call APIs directly, Shippy routes all data queries through a purpose-built CLI that wraps the Skylight API. This ensures deterministic, auditable tool calls. Skills include querying vessel events, looking up exclusive economic zones (EEZs) and marine protected areas (MPAs), interpreting track data, and generating interactive map links. Analysts can verify every answer through deep links back to the source map. The soul explicitly forbids legal determinations or speculation beyond available data.
- Shippy uses a three-part architecture: soul (behavioral prompt), skills (versioned markdown tool specs), and config (runtime settings like LLM and harness).
- Instead of raw API calls, Shippy talks to Skylight via a deterministic CLI, making tool usage predictable despite the nondeterministic LLM.
- Skills follow the same markdown spec as Claude Code and Codex, making them easy to write, version, and revise.
Why It Matters
Shippy shows how to build reliable AI agents for critical domains where mistakes have real costs.