Orchestrator Tools (The Internal Cascade)

The Orchestrator has direct access to a cascade of internal tools that Research Agents do not. These are the Orchestrator's eyes into the member and the trip. They give it the context it needs to scope and dispatch jobs without leaking member state to workers.

Why These Live on the Orchestrator

• Workers stay stateless and domain-pure
• Member data is gated behind one boundary
• Curation is a product decision the Orchestrator enforces; workers can't bypass it
• Easier to audit, log, and version the "what does the orchestrator know" surface as one unit

The contract is: the worker receives a pre-narrowed brief. It does not have access to the cascade.

Tier 1: Member & Trip Context

Tool	Returns
get_member_profile	Member identity, tier, preferences, loyalty status, dietary restrictions, accessibility needs
get_trip	Trip dates, destination(s), companions, budget envelope, itinerary state
get_trip_history	Past trips, prior reservations, feedback, no-go lists
get_member_preferences	Cuisine likes/dislikes, price comfort zone, party-size patterns

These are the first calls the Orchestrator makes on any turn. They establish the working context.

Tier 2: Destination & Knowledge Tools

Tool	Returns
get_destination_info	Curated info for a city/region: neighborhoods, transit overview, cultural notes
get_seasonal_context	What's in season, local events, holidays that may affect availability
get_currency_and_payment	FX rates, payment norms, tipping expectations

These let the Orchestrator enrich briefs with situational context the worker would otherwise have to discover or guess at.

Tier 3: Internal Curation ("Known Good" Lists)

Tool	Returns
get_curated_dining_list	Concierge-vetted restaurants per destination, with tags (signature, hidden gem, splurge, casual)
get_blacklist	Members or trips flagged for known issues (past complaints, dietary emergencies)
get_partner_inventory	Real-time availability/holds from partner systems (OpenTable, SevenRooms, direct hotel dining, etc.)

These are the "house advantage". The Orchestrator can pre-filter against trusted, curated data before dispatching a worker. A worker brief often starts from a curated list, narrowed by the Orchestrator, rather than an open-ended search.

Cascade Flow

Example Brief

{
  "task": "Find 3 dinner reservations in Tokyo for Feb 13-17, party of 2",
  "constraints": {
    "cuisines": ["sushi", "kaiseki", "izakaya"],
    "max_price_per_person": 200,
    "neighborhoods": ["Ginza", "Shibuya", "Roppongi"],
    "dietary": ["no_pork"]
  },
  "curated_shortlist": ["sukiyabashi-jiro", "den", "kyubey"]
}

The curated_shortlist field is the Orchestrator's pre-filtered list. The worker searches within known-good inventory rather than the open web.