The 8 AI agents inside the IMPT swarm
Deep guide to all 8 agents: hotel search, flight advisor, gift concierge, franchise, country owner, B2B marketing, dashboard, portal router.
"AI swarm" is a strong claim, so we'll be specific. The IMPT system is eight agents, each with a single job, communicating over a shared task bus. Here is what each one actually does.
1. Hotel-search agent
The big one. Indexes 1.7M hotel URLs across 195 countries, queries live partner rate feeds at the moment you ask, ranks results against the full prompt — geography, amenities, star rating, price band, family or business profile. Returns a shortlist in 6-9 seconds. The other agents enrich its output.
2. Flight advisor
When your prompt implies travel from somewhere — "trip from Dublin to Lisbon, last week of October" — the flight advisor scores route options against time, total cost, and emissions, then bundles the cheapest realistic flight + hotel pairing into the recommendation. It does not book flights yet (we route you to the airline direct), but it shapes the hotel night-count and arrival timing.
3. Gift concierge
Hotel-stay add-ons. Champagne, room flowers, restaurant booking, spa slot, airport transfer. The concierge agent talks to the hotel's concierge desk via partner APIs and arranges the extras as part of the same booking. Most useful for anniversaries, surprise birthdays, and corporate client trips.
4. Franchise agent
A B2B layer. Franchisees and small hotel groups onboard their inventory through the franchise agent — branded landing pages, white-label search, commission flow back to the IMPT network. Travellers don't see this agent directly, but it is why the swarm covers boutique chains other OTAs miss.
5. Country owner agent
Each country has a designated owner agent that maintains local context — currency quirks, public holidays that affect rates, regional brand presence, language nuance for the hotel's guest-facing copy. When you ask "best 4-star in Lisbon for Christmas week", the country owner agent for Portugal is the one telling the hotel-search agent that the second week of December is rate-stable while the third week spikes 40%.
6. B2B marketing agent
Drives the corporate channel. ESG-ready packages, scope-3 emissions reports, downloadable invoice bundles for accounting. Travel managers interact with this agent to spin up a private travel programme for their company's bookers without leaving the swarm.
7. Dashboard agent
Your account view. Past bookings, on-chain offset hashes, price-watches, saved hotels, IMPT token balance and the IMPT Visa card statement. Everything personal lives behind this agent and only this agent.
8. Portal router
The traffic conductor. Decides which of the other seven agents to call, in what order, and which one writes back to you. You don't see the router directly — but it is why a complex query that touches geography, amenities, gifts, and ESG reporting can be answered in one chat turn instead of 14.
Why eight, not one?
A single monolithic LLM tries to be good at everything and ends up mediocre at most things. Eight specialists each tuned to their layer — geography, ESG, payments, concierge — cost less per query, scale independently, and let us swap models per agent when a better one ships. The user just sees one chat. The swarm does the rest.
Read next
- AI hotel search vs traditional booking
- For corporate travel managers
- How the carbon offset works
- AI hotel search in Dublin
- AI hotel search in Tokyo
Try the swarm
Two ways in. Ask the AI in plain English, or jump straight into the booking flow with your dates already set.
Frequently asked
Are all 8 agents always called?
No. The portal router only invokes the agents your query needs. A simple 'Paris next weekend' query touches three agents; a corporate ESG query for a 14-city tour touches all eight.
Can I talk to a specific agent directly?
Yes — power users can scope a query, e.g. 'b2b agent: prepare a scope-3 report for Q1'. Most travellers just type natural language and let the router do the work.
Are the agents the same model under the hood?
No. We pick the best model per layer. Heavier reasoning agents like the country owner run on the strongest models; the dashboard agent uses smaller, faster ones.
Will there be more agents?
Yes. A loyalty agent and a multi-city itinerary agent are in development. We will name them and ship them rather than silently expanding existing ones.