A copilot waits for your question. An agent doesn't. Give an agentic system a goal — "monitor the gensets," "keep this twin in sync," "plan tomorrow's worklist" — and it decomposes the task, picks its tools, calls them, reads the results, and acts. That step from answering to deciding what to do next is what "agentic" actually means, and it's the reason the word matters aboard a ship more than in a chat window. Maritime-GPT analysis
Key findings
- Peer-reviewed maritime agentic systems exist now — an AAAI-published shipping digital-twin agent and an Ocean Engineering vessel-condition agent are the two cleanest examples. Research demonstration
- The winning pattern: the LLM orchestrates tools, it doesn't compute — anomaly detection, retrieval, and metrics stay in deterministic code the agent invokes. Research demonstration
- Agentic ≠ autonomous. The published systems act on information workflows — queries, monitoring, explanation — not on propulsion or steering. Established
- The authority boundary is the hard problem: an agent that can act is an agent that can act wrongly, and maritime law assigns the consequence to a human. Maritime-GPT analysis
- The MASS Code gives the frame — goal-based, risk-assessed, master-retains-responsibility — but agentic software was largely designed after it was drafted. Maritime-GPT analysis
What the first systems actually look like
The IBM/AAAI shipping agent puts a reasoning-and-planning LLM in front of a vessel digital twin. Ask it a question — in a fleet manager's, master's or engineer's persona — and it decomposes the request, selects specialised tools, queries the twin, and synthesises an answer. The demonstration's point is architectural: language becomes the universal interface to a complex system, and the agent handles the navigation of it. Research demonstration
The vessel-condition agent goes further into operations: a Transformer autoencoder watches high-frequency sensor data and flags anomalies; an LLM agent then investigates — invoking domain-specific tools to pull measurements, compute derived metrics and query the anomaly outputs — before writing up what it found and why it matters. Validated on 91 days of real offshore-vessel data, it detected all four test fault scenarios with lead times of 97 to 637 minutes. Field trial
Notice what both designs share: the language model never touches the raw numbers directly. It calls tools that do. That constraint — deliberate in both papers — is the difference between an agent and a very articulate random-number generator. Maritime-GPT analysis
The agentic stack, maritime edition
Maritime-GPT synthesis of the pattern common to the published maritime agentic systems. Maritime-GPT analysis
The authority boundary is the whole game
Enterprise agentic AI fails loudly but recoverably — a bad tool call gets retried, a wrong answer gets flagged. Shipboard agentic AI fails differently: a confident plan acted on without a checkpoint can move steel. The MASS Code resolves this at the regulatory level — the master retains responsibility at all times — but that resolvesliability, not architecture. The engineering question the papers leave open is where the agent's agency ends: which tool calls are read-only, which write, which require a human confirmation, and how that's enforced rather than suggested. Maritime-GPT analysis
The honest current answer from the literature: read-mostly. The published systems observe, retrieve, compute and explain. None of them actuate. That isn't an accident — it's the boundary that makes them publishable, and it's the boundary a deployable system must make load-bearing, in code, not convention. Maritime-GPT analysis
What we know / don't / being tested
What we know
- Planning + tool-use agents work over real maritime data and workflows.
- Tool-mediation suppresses hallucination far better than letting the model compute directly.
- Mid-size agents run on shipboard edge hardware — this isn't cloud-locked.
What we don't know
- Behaviour under goal conflict — when a plan's subtasks point at contradictory actions.
- Multi-agent interaction — ship agents reasoning about shore agents about other ship agents.
- Long-horizon reliability — published runs are demonstrations and short validations, not months of autonomy.
Being tested
- Agent interfaces to digital twins (AAAI) and condition monitoring (Ocean Eng).
- Function-calling on edge hardware with domain-constrained tool sets.
- Persona-aware interfaces — same agent, different views for master vs engineer vs fleet ops.
Maritime-GPT analysis
The near-term agentic ship is a better-run ship, not an unmanned one. The defensible deployment today is agents on information workflows — watch summaries, maintenance triage, document retrieval, twin interrogation — where acting meansinforming. Crossing into actuation belongs to a different assurance regime entirely, and the research correctly hasn't crossed it. Shipowners should evaluate "agentic" claims by one question: what can it do, and what's between that and the steel? Maritime-GPT analysis
Open research problems
- Enforced authority boundaries — hard, verifiable limits on which actions an agent may take, not prompt-level suggestions.
- Agent observability — logging and replay of tool-call sequences for incident investigation under the SMS.
- Certification of open-ended planners — how a goal-based MASS risk assessment covers software that chooses its own subtasks.
- Safe stopping — documented agent behaviour on tool failure, contradictory evidence or mid-plan context loss.
Evidence scorecard
| Dimension | Rating | Basis |
|---|---|---|
| Evidence strength | Moderate | Peer-reviewed demonstrations incl. one on real vessel data. |
| Research maturity | Experimental | Working prototypes; none deployed in operational loops. |
| Maritime deployment | None verified | No published operational agentic system actuating aboard. |
| Confidence in conclusions | High | Architecture claims trace to named papers; boundary analysis labelled. |
Related tools & reading
- Agentic AI for marine engineering diagnostics
- LLMs + digital twins
- Navigation-GPT architecture
- The MASS Code
Sources
- Agentic AI for Digital Twin — Timms, Langbridge, Antonopoulos, Mygiakis, Voulgari, O'Donncha (IBM Research)Peer-reviewed demo (AAAI 2025)
A shipping digital twin fronted by a reasoning-and-planning agent that decomposes tasks and coordinates specialised tools — personas from fleet owner to technician. The cleanest published maritime agentic architecture. - Agentic AI for autonomous condition monitoring and predictive maintenance of marine vessels — Ocean Engineering 364, art. 127078 (2026)Peer-reviewed (2026)
An agentic monitor — Transformer autoencoder detects anomalies, an LLM invokes domain tools to explain them; validated on 91 days of real offshore-vessel data, 4/4 fault scenarios caught with 97–637 min lead. - Deploying Efficient LLM Agents on MASS: Fine-Tuning, RAG, and Function Calling (MARTIAN) — Information (MDPI), 17(3):284 (2026)Peer-reviewed
A function-calling maritime agent engineered for edge hardware — evidence that agentic patterns don't require cloud scale. - IMO adopts first global Code for autonomous ships — International Maritime OrganizationPrimary — regulation
The frame all agentic ship software lands in: risk assessment, human accountability, and the master's retained responsibility.
System behaviours quoted from the cited papers; regulatory framing from IMO. Accessed 5 October 2026. Editorial analysis, not engineering or legal advice.