Five customer-facing agents plus a personal one. Each agent calls the calculators as tools and writes the prose on top, with citations. The LLM never produces a number that drives a decision — that wall is enforced by CI.
Each agent is a tool-use loop, not a sampling machine. The tools do the math; the agent writes the explanation.
Routes regulatory questions through the hybrid RAG corpus and the compliance calculators. Surfaces the regime, the article, the implementing dates and the carve-outs — with chunk citations on every claim. Uses the live customs and sanctions datasets as tools, never as paraphrase.
Reasons over the eight commodity briefs and the six Asia origins. Calls the preferential-origin engine to surface framework qualification, the trade-defence database to flag stacked duties, and the sanctions screen against beneficial owners before recommending a shortlist.
Knows the lane shape across thirty origin × destination combinations, the bonded options at six EU hubs, and the transit windows by carrier class. Calls the freight calculator and the customs-procedure engine, then surfaces the lane that fits the cargo.
Reasons over the calculator-grounded landed-cost math, the FX hedge windows from the live market, and the working-capital cycle for the lane. Surfaces the cash-locked window in days and the basis-point sensitivity in euros.
A meta-agent that decomposes the question into specialist sub-questions, delegates to compliance, sourcing, logistics or finance as needed, and merges the findings into a single calculator-grounded answer. Twenty-seven tools across the four specialists plus its own routing tools.
A separate agent confined to the signed-in user’s saved plans, portfolios, monitoring alerts, compliance deadlines and agent memory. Eight personal tools — recallMemory, rememberForUser, forgetForUser, getMyComplianceDeadlines, plus the orchestrator’s read-only set.
Every agent calls calculators as tools. The Anthropic SDK is import-banned in any module that touches money math — the ban is checked by CI on every commit.
Every regulatory claim cites the chunk identifier it came from, with a confidence tier. Every plan is stamped with the calculator version, the data-snapshot date and the customs mode used — reproducible on any past date.
Every irreversible action (delete, surrender, file) calls requestHumanReview before executing. The agent recommends; the human commits.
Available behind sign-in at the Ask-the-agent surface in the cockpit. Same calculators, same RAG corpus, same audit chain — for your saved plans and portfolios.