Section 01
The call on "agents competing"
The brief says agents compete. It doesn't say what they compete over. There are three plausible readings — and the product looks completely different depending on which one you pick.
- Auction — agents bid against each other for scarce inventory. That's Priceline with an AI skin.
- Assistant — one user's agent searches OTAs harder than the user would. Useful, but Booking, Expedia, or ChatGPT can copy it in a quarter.
- Scout market — personal scouts collaborate and compete inside city-level deal quests, with roles, reputation, and a mandate. Together they surface price anomalies, validate tradeoffs, and negotiate supply.
Section 02
The user — a flexible hunter
Picture them: 25 to 40, comfortable spending on travel, but emotionally motivated by beating the market. They book two to four trips a year, forward Going.com emails to friends, compare neighborhoods obsessively, and love the story of finding a better deal.
They aren't chasing the cheapest hotel. They want a deal that feels earned: "I got something better than the obvious search result because I played the hunt well."
Section 03
The scout relationship
Every user has a Scout: a persistent AI character that represents their taste and their constraints in the market. The scout has visible strengths and a personality — light enough to feel like a real character, but never heavy enough that booking starts to feel like gambling.
Product rule: one user, one Scout. The strengths below describe that representative's behavior; they are not separate agents the user has to manage. When a trip has multiple travelers, their Scouts form one shared party around the group mandate.
What the user actually configures. Stats describe the scout, but the user still has to brief it. Anthropic's Project Deal experiment surfaced four useful knobs — mandate, negotiation style, standing rules, and persona — and onboarding hits all four in a quick chat, not a settings page:
Honest read of the Project Deal results: scouts followed all four knobs reliably, but only mandate clarity and model quality actually moved the final price. Style and persona were felt by the user, not measured in the outcome. So I'm treating mandate and standing rules as first-class config, and style and persona as character — load-bearing for trust, not for price.
Net effect: the agent feels like a representative, not a chatbot. Users don't search — they brief a scout and send it into a quest.
Section 04
Core loop
Hotel booking is episodic. But scouts can stay useful between trips. So the loop runs on two cadences — a daily free patrol, and a longer trip-mode hunt.
Section 05
Monetization
I'm keeping the small fee from the brief, but reframing what it buys. Not a per-search charge — a stake on a hunt.
Paid stat rerolls, loot boxes, and pay-to-win scout upgrades are a hard no. The product can feel game-like in habit, but the economics around real money have to stay boring and trustworthy.
Section 06
Negotiation realism
Worth being precise about what "negotiation" means on day one. It's not the scout phoning every hotel.
Together those four levers keep day-one negotiation real but bounded. The user still feels represented by an agent — and the product doesn't promise autonomous hotel-by-hotel negotiation before there are enough hotels in any one city willing to take that call (the supply-density problem).
Section 07
Retention — the daily free action
The daily free action is Send Scout on Patrol.
A patrol can:
- Judge 3 live or simulated hotel deals: take, pass, watch, explain why
- Validate a public quest lead: good location, fee trap, fake discount
- Collect city intel: "Shibuya weekends are softening" or "Príncipe Real boutiques drop Thursdays"
- Improve scout memory and lightweight stats
- Earn reputation for useful signal
Stronger than a passive deal feed, because the user has a role. They're training an agent and contributing to a market — not scrolling travel content.
Section 08
Social mechanic — organic growth
The social mechanic is Party Quests. Users invite friends or compatible strangers because scout composition matters. For group travel, the trip runs as one shared quest: each traveler brings their one Scout, and the party resolves price, location, flexibility, and review risk against a single group mandate.
- Quest invites — "Bring your Radar Scout; ours keeps missing hidden fees."
- Live quest follows — friends can watch a hunt unfold without booking.
- Quest recap cards — "4 users, 4 Scouts, 9 days, 37 passes, $340 saved."
- Scout reputation — useful scouts become recognizable inside city quests.
The shareable object isn't a coupon. It's the story of a hunt.
Section 09
Positioning + category
Not: an AI travel agent, an opaque booking site, or a cheaper Priceline.
The category claim: multiplayer travel scouting. Users send personal AI scouts into a live market, where deal intelligence is discovered, validated, and acted on — socially.
Reference points: Going.com proves the deal-hunter segment exists and pays. Pokémon Go and MMO-style party quests show that location-based daily activity works at scale. Discord and Strava show that people will form social identity around a hobby. Priceline and Hotwire are what this explicitly isn't.
Section 10
Calls made on ambiguous parts — explicit
| Call | Reasoning | Trade-offs |
|---|---|---|
| Agents compete and cooperate inside quests | This makes the agent mechanic user-visible and avoids a thin "AI searches harder" product. | Heavier UX surface and onboarding than a single-agent chat; risk of feeling busy to a casual booker. |
| User = flexible hunter | Broad travelers won't tolerate surprise or tradeoff; hunters will. | Narrower addressable market on Day 1; cedes the bulk of OTA volume to incumbents until the niche compounds. |
| Scout = persistent representative | The agent has memory, strengths, and reputation, not just a chat window. | More state to engineer and protect; if users don't bond with their scout, the persistence buys nothing. |
| Daily retention = patrols, not daily booking | Travel is too low-frequency for a booking-only habit. | Patrols burn compute and design effort without producing direct revenue; must stay cheap and feel useful. |
| Discount promise = benchmarked savings | A qualified strike must beat a transparent all-in benchmark or the hunt stake rolls forward/refunds. | Caps revenue on hunts that don't qualify; benchmark sourcing must be defensible or the guarantee becomes a liability. |
| Social growth = party composition | Invites are useful because different users' Scouts improve quest outcomes. | Slower than referral bounties; only works if quests are interesting enough that inviting friends feels worth it. |
| Small fee = hunt stake | Preserves the brief's paid primitive, but frames it as commissioning work rather than paying for a search result. | Friction at the top of funnel; some users won't pay before seeing value, so onboarding has to earn the first stake. |
| Game layer stops before checkout | Discovery can feel playful; booking must feel serious, transparent, and reversible where possible. | Loses some emotional payoff at the moment of purchase; the handoff between fun and serious modes is hard to design cleanly. |
Section 11
What I'd pressure-test with another week
- Does the game layer cheapen trust? The answer is probably visual restraint — map, roles, reputation, progress. No loot-box energy anywhere near a $500 booking.
- Will public quests get enough shared intent? Start in dense destinations with broad quests ("Tokyo weekend under $180") before narrowing the targeting.
- How real can negotiation actually get? Day-one negotiation is bounded but real — better private rates, fee waivers, term concessions, timing windows, and group-demand signals. Direct hotel deals scale city by city, after demand concentrates.
- Can scout stats be more than cosmetic? They should shape ranking, explanation, and matching — but never create paid access to inventory.
- Where do operational edge cases break this? Booking failures, refunds, resort-fee disputes, and overbookings will all need human-assisted support before users trust autonomous checkout.