Seven agents and an assistant.

An intelligence layer coordinates the agents through one evidence model: it defines the decision, horizon, grain and constraints; registers the expected evidence; normalizes shared identities; runs reconciliations before any modeling; keeps source facts, accounting interpretations, estimates, forecasts and recommendations in separate columns; and evaluates recommendations jointly, because pricing, promotions, labor, purchasing, quality, demand and working capital interact. It presents a frontier across profit, cash, service, risk, guest experience, food safety and implementation complexity, and routes every consequential write to approval.

Proposals. Every recommendation is a typed proposal (vendor order, schedule band, reclassification, data confirmation, graph correction, note, experiment, cash action, journal approval) with a risk class (read, propose, act), a policy (auto, approve, owner-only), the nodes it affects, the evidence it cites, the run and snapshot it came from, and a required expected-effect range before it can be shown. Proposals are immutable, tenant-scoped and idempotent; at most 12 a day per restaurant.

Approvals. Your decisions are approve, reject, edit or defer, recorded with the note, the edit diff, dwell time and the rank at which the proposal was surfaced. Routine supported operations run under standing, versioned, evidence-based policy. Posting journals, paying bills, moving money, changing vendor banking, placing orders, changing approved substitutions, filing taxes and acting on employee allegations always require a person. Changed allergens or units never pass the routine path.

Receipts. Every external-facing intent gets a sha256 idempotency key. Receipts are write-once, never rewritten, never deleted; a pending marker survives a crash so nothing is retried blindly. Every executed operation produces an execution receipt binding the operation, work item, decision, authorization, actor, output hash and measured resource use. Delivered work, an achieved objective and evidence that an intervention happened are three different records.

The assistant is a durable state graph (scope, plan, tool, evaluate, review, clarify, critic, finalize) with real conditional execution and durable interrupts, so an investigation can pause for your answer and resume with its evidence intact. A planner selects one allowlisted tool per step, and a separate critic checks the draft against the tool evidence; deterministic gates require valid citations before any answer can finish. Budgets: 16 planning steps, 14 model calls, 200,000 tokens, a 45-second call timeout, three failed attempts, three critic rounds. Conversation history, up to twelve prior turns, is encrypted, and every new turn rechecks the facts.

• It can investigate: where to fix food-cost visibility first; whether the sales records reconcile; what a supplier price change costs per month and which dishes it touches; how much cash the next 13 weeks need; whether a payroll correction replays cleanly; what to prep for Friday given the range.
• Its tools: source coverage; monthly P&L and balance-sheet drafts; reconciliation differences; statement cash observations; POS sales coverage; inventory counts; saved forecast evaluations; recipe dependencies and recipe-based ingredient demand; an evidence catalog with bounded reads; a Decimal calculation sheet with units, assumptions and formula receipts; work-item recording; a live seven-day forecast with conformal intervals and a chart; operating status and playbook proposals; advertising guidance and scoring.
• Boundaries: raw bank, guest and employee rows never reach the model; calls carry approved aggregates only; accepting a draft journal means “accept for CPA review”, never posting.
• Proof: tested against independently authored owner tasks, every reserved task fully or partially completed with zero critical failures across 39 recorded episodes.