Inside Shippy: How an Ocean-Monitoring AI Agent Was Built to Be Trusted, Not Just Smart
The team behind Skylight's maritime AI explains why reliability, not raw intelligence, was the hardest engineering problem they faced.
Key points
- Shippy is an AI agent, software that can carry out multi-step tasks on its own, built to help maritime analysts track illegal fishing and vessel activity in real time.
- Shippy currently runs on Claude Opus 4.6, a large language model made by Anthropic, the technology behind many advanced AI chatbots.
- Every user gets their own isolated session: no data from a fisheries officer in the Philippines ever touches data belonging to another government agency.
- The team built a custom command-line tool, a simple text-based interface, to stop the AI from making subtle errors when pulling live data from satellites and vessel signals.
- Skylight serves hundreds of government agencies and NGOs across more than 70 countries.
If a patrol vessel gets sent in the wrong direction because an AI gave a bad answer, real people in dangerous jobs bear the cost. That single fact shaped every design decision behind Shippy, the AI assistant built into Skylight, a platform for tracking ships at sea.
The team published a detailed breakdown of their approach on Hugging Face. What they describe is less a story about a clever AI and more a story about building guardrails around one.
What actually stops an AI agent from getting things wrong?
The answer, the Skylight engineers found, is layers. Not a smarter model, but a series of predictable, testable steps that narrow what the AI can mess up at each stage.
Shippy is built around three concepts the team calls soul, skills, and config. The soul is a set of written instructions baked into the system telling Shippy what it is and what it will not do. It will not decide whether a vessel is breaking the law. It will not guess beyond what the data shows. Those limits are written out explicitly, which means a human can read and change them.
Skills are separate instruction files that handle specific tasks: looking up a vessel's position, finding the boundary of a country's Exclusive Economic Zone (the sea area where a nation controls fishing and resources), or generating a map link so an analyst can verify the answer themselves.
Config covers practical settings: which AI model to use, which software framework to run it on. Swapping to a different model is a one-line change, not a rebuild.
The trickiest part was data. Skylight's API, the connection point between Shippy and live satellite data, handles dozens of input types, complex filters, and pagination, meaning it delivers results in chunks rather than all at once. Early versions of Shippy built its own data requests from scratch. The results looked correct but were quietly wrong: missing records, garbled location data, misread filters.
The fix was a custom command-line tool the team calls the Skylight CLI. Instead of building a raw request, Shippy issues one simple command and the CLI handles all the complexity underneath. Each layer of the system can be tested on its own. If something breaks, engineers know exactly where to look.
Isolation was equally critical. When a fisheries officer logs in and asks Shippy a question, their data and conversation history must never be visible to any other user. The team built a hosting system called Mothership that spins up a fresh, contained environment for every single user session. When the conversation ends, that environment disappears.
For ordinary people, the practical takeaway is this: the design choices that make Shippy trustworthy in a high-stakes government setting are the same ones that matter whenever AI touches sensitive decisions. Explicit limits, testable steps, and strict data separation are not optional extras.
