The Allen Institute for AI (Ai2) has disclosed the underlying architecture behind Shippy, a high-stakes maritime AI agent deployed across more than 300 partner organizations in 70 countries. Built by the Skylight team, the system relies on Anthropic's Claude Opus 4.6 and the open-source OpenClaw agent framework to provide real-time ocean surveillance and fisheries enforcement support. By decoupling deterministic software tools from nondeterministic language models, Ai2 aims to eliminate dangerous hallucinations in critical operational environments.
A Modular Three-Part Architecture
To build an agent capable of assisting fisheries officers and ocean analysts without risking resource misallocation, Ai2 structured Shippy into three distinct layers: a soul, skills, and runtime configuration. The soul consists of a system prompt defining Shippy's persona and non-negotiable behavioral boundaries, such as refusing to make legal determinations or military intelligence assessments. Skills are written as human-readable Markdown files matching the skill spec used by tools like Claude Code and Codex, detailing standard operational workflows like resolving Exclusive Economic Zone (EEZ) boundaries.
Together, the soul and skills are packaged into a versioned Docker image. The runtime config manages settings like API credentials, agent harnesses, and model choice, allowing developers to swap underlying models without rebuilding the core agent container.
Key Architectural Features
- Deterministic CLI Layer: Instead of letting the LLM generate raw API queries with complex nested filters, Shippy executes commands through a dedicated Skylight command-line interface (CLI) that writes JSON responses straight to disk to prevent pipe buffer failures.
- Isolated Pod Hosting: Ai2 engineered Mothership, a dedicated Kubernetes platform that provisions an isolated, ephemeral sandbox per user session, injecting individual user credentials so proprietary watchlist data is never shared across accounts.
- Integrated Ocean Models: Shippy calls specialized sub-models, such as Ai2's Atlantes model for vessel behavior classification and ProtectedSeas databases for Marine Protected Area (MPA) boundaries.
- Scenario-Based Evals: Using the open Harbor framework, Ai2 evaluates the combined agent stack against live streaming satellite data, employing LLM judges that score responses on weighted criteria written by domain experts.
Live Evaluation and Operational Limitations
Because static benchmarks fail to measure real-world operational safety, Ai2 runs continuous evaluations against streaming satellite and vessel broadcast data. In recent test runs, the team reported strong performance on source attribution and data guardrails, but highlighted notable operational edge cases that required further refinements.