How it works

When a container sits longer than its free time, the carrier charges demurrage and detention by the day, per container. The rates are published, the rules about how they can be billed are federal law, and yet the moment a vessel is actually delayed the response is a scramble: someone digs out the tariff PDF, someone else guesses at the reroute cost, someone emails the carrier, and the clock runs the whole time.

The information needed to decide well is public. It's just scattered across carrier tariff documents, port advisories, regulatory filings and freight indices — and nobody has time to gather it while the meter is running.

Plan. Describe a shipment and a pipeline of agents researches it against the live web: freight rates, tariffs and duty, port congestion, weather, geopolitical and supplier risk. You get a risk score, a cost and delay forecast, a recommended entry port, and a dated action plan — every figure carrying the source it came from.

Track. Keep a shipment and it stops being a one-off report. The pages that determine what a disruption would cost on that lane — the port's advisories, the carrier's demurrage tariff, the FMC billing rule — are put under scheduled monitoring, and changes land on the shipment.

Act. If a vessel is actually disrupted at a monitored port, a room opens. Seven agents across three frameworks negotiate recovery options costed against the real published tariff, vote to a quorum, and an adversarial agent argues against the front-runner. Then a human approves or rejects, and the whole exchange becomes an audit dossier.

The join is the tracked shipment. Planning produces one; monitoring watches its lane; an incident at its entry port attaches to it. That's why the watchlist is the home page — it's the list of things the rest of the product is about.

The attribution is deliberately conservative. A shipment is linked to an incident only when its resolved entry port is the same monitored port, and the link stores the sentence explaining why. It's labelled a port-level match, because that's all it is: it does not claim your cargo is aboard the affected vessel, and the data available couldn't support that claim.

Live web access runs through Anakin. Models reason; they do not browse. Four capabilities, used in specific places:

• Search (/v1/search) — every planning agent: product, commodity, freight, port, weather, geopolitics, supplier, regulation, the port recommender, duty documents, and driver prices. About 25–30 cited searches per analysis. That is the transparency panel on a shipment page.
• AI structured extraction (/v1/url-scraper/scrape) — ops: the carrier demurrage PDF becomes per-day tariff rows the room negotiates over, and a public BoL page is the importer fallback (always labelled inferred). Pages that 403 a plain fetch (Port of Los Angeles notices) still yield dated text here.
• Scheduled monitors (/v1/monitors) — when you track a shipment, the port advisories, carrier tariff, and 46 CFR Part 541 billing rule go on a watch. Changes land as web signals with the source URL. HMAC-signed webhook; cost scales with distinct pages, not with how many shipments share them.
• Deep research (/v1/agentic-search) — implemented, not used in Plan. It returns no citation list, so it is never the basis for a number this product asserts.

Extraction is how a number becomes citable rather than estimated. The rule the whole product is built on: nothing is invented to fill a gap. If a source can't be reached, the figure is absent and labelled, not estimated into place. Wire was evaluated and is unused — no bill-of-lading or customs action in that catalog.

Three different frameworks on purpose, and the dissenting agent runs on a different model from the ones it argues with — so the objection isn't one model agreeing with itself.

• LangGraphSentinel — Watches AIS, opens the room, drives the phases.
• LangGraphLogistics — Proposes reroutes, holds and expedites.
• PydanticAICarrier — Negotiates from the published tariff and FMC billing rules.
• PydanticAIFinance — Costs each option against real demurrage exposure.
• CrewAIProcurement — Looks for alternate supply and air-freight cover.
• PydanticAICustomer impact — Assesses downstream SLA and order risk.
• PydanticAIDissent — Argues against the leading option, on purpose.

1. 1DetectionA Sentinel agent watching the live AIS feed spots a vessel dwelling in a monitored anchorage.
2. 2NegotiationIt opens a room and recruits the other agents, who start proposing and challenging recovery options.
3. 3OptionsEach proposal is costed against the carrier's published demurrage tariff, not an estimate.
4. 4EvaluatingThe agents vote to a quorum, and an adversarial Dissent agent argues against the front-runner.
5. 5Approval gateA human reads the recommendation and the objection, then approves or rejects it. Nothing acts on its own.
6. 6ApprovedThe decision and its reasoning are recorded in an audit dossier.

Because it's real. It reflects an actual AIS feed watching three port complexes — Los Angeles / Long Beach, New York / New Jersey and Singapore — and it lights up when a vessel is genuinely dwelling where it shouldn't. It isn't a scripted demo that plays on load. The vessel map is live either way, so you can see the feed working even when nothing is wrong.