From Research to Reservation: A Companion Pattern for Evidence-Carrying Travel Agents
A travel agent becomes dependable when a recommendation remains traceable as it moves from search result to itinerary—and stops short of booking until its assumptions are confirmed.
This article extends Install Once, Drain Anywhere: Building a Travel Agent That Survives Real Trips with a narrower angle: how to preserve evidence while turning flexible travel research into a reviewable booking decision.
The gap between a good option and a bookable option
Travel research often begins with an open-ended request: find a practical route, keep the cost reasonable, avoid an overnight connection, and make the plan fit a person’s schedule. Search tools can return plenty of plausible options. The difficult work starts afterward.
A recommendation is not yet a booking candidate merely because it has a price and a departure time. The agent must know what the price includes, whether the date and airport are correct, how long the connection lasts, and whether the source is current enough to support an action. A result can be useful for exploration while still being unsafe to purchase.
That distinction suggests two separate outputs. First, produce a research set containing alternatives and evidence. Then, after constraints are resolved, produce a short booking set whose claims are specific enough for a person to approve.
Preserve the evidence at the moment of discovery
The most reliable place to capture evidence is immediately after a result is found, not at the end of the itinerary-writing step. For each candidate, retain at least:
- the source URL or provider identifier;
- the retrieval time;
- the exact departure and arrival airports;
- local departure and arrival times, including the dates;
- carrier, flight or train numbers, and connection points;
- displayed price and currency;
- baggage, fare, cancellation, and change conditions when available;
- assumptions or fields that still need confirmation.
This does not require saving an entire website. It requires preserving enough context to answer a later question: “Why did the agent recommend this?” A bare link is weak evidence because prices change, sessions expire, and search pages may no longer show the same inventory.
The evidence record should also distinguish observed facts from derived values. “The provider displayed a 90-minute connection” is different from “the connection is comfortable.” The first is a source-backed observation; the second is a judgment that should be explained against the traveler’s preferences and risk tolerance.
Turn preferences into decision rules
A travel request usually contains more than one kind of constraint. Some are hard boundaries: a traveler cannot depart before a certain time, needs an accessible airport, or must arrive before an event. Others are preferences: fewer stops, a favorite carrier, or a willingness to pay more for flexibility.
An agent should represent those categories separately. Hard constraints filter candidates. Soft preferences rank the survivors. If everything is blended into one opaque score, the itinerary may look polished while hiding a disqualifying tradeoff.
A compact decision record might say:
``text hard: departure_after: 08:00 local arrival_before: 21:00 local max_connections: 1 prefer: refundable: true shorter_total_duration: true accept: higher_price_for_refundability: true ``
The exact syntax is less important than the behavior. When no candidate satisfies every hard constraint, the agent should report the conflict and ask which constraint may be relaxed. It should not silently substitute a red-eye, a different airport, or a nonrefundable fare merely to complete the itinerary.
Build an itinerary as a chain of commitments
An itinerary is more than a list of segments. It is a chain of commitments with different levels of reversibility.
Research commits the agent only to describing an option. Selection commits it to recommending one option over others. Booking commits money and creates an external reservation. Those transitions should not be collapsed into one tool loop.
A useful itinerary can therefore label each item:
- researched: evidence captured, no selection made;
- recommended: preferred candidate identified, assumptions exposed;
- approved: traveler accepted the candidate and total cost;
- booked: provider returned a confirmed reservation;
- needs review: evidence is stale, contradictory, or incomplete.
This vocabulary protects both the traveler and the operator. A generated itinerary may contain a “recommended” hotel without implying that a room is held. A booking confirmation should include the provider’s reservation code and the exact passenger or guest details used—not merely a sentence saying that booking succeeded.
Use verification checkpoints before external effects
The booking boundary deserves a deliberate checkpoint. Before sending payment or confirming a reservation, the agent should restate the material facts in a compact review surface: traveler, dates, origin and destination, segment times, fare conditions, total price, currency, and any important assumptions.
The checkpoint should also identify what will happen next. If the action is approved, the agent will place the reservation. If the traveler declines or evidence has become stale, the agent will return to research. This makes the external effect explicit rather than burying it inside a larger autonomous plan.
After booking, verify the result from the provider response. Capture the confirmation number, ticket or reservation status, purchased fare, and cancellation deadline where available. If the provider times out after payment may have been submitted, mark the state as uncertain and reconcile with the provider before retrying. “Retry until a confirmation appears” is how duplicate reservations happen.
Keep the itinerary useful after the booking
Evidence should not disappear once the reservation is made. The final itinerary should preserve the relationship between the booked record and the original decision: which option was selected, when it was checked, what was paid, and which conditions apply.
That record supports ordinary travel needs—finding a confirmation, checking a change rule, or understanding a connection—but it also makes future agent assistance safer. A later request to modify the trip can begin from verified reservation facts instead of treating an old conversational summary as authoritative.
The same structure works for non-flight travel. Rail tickets, lodging, car rentals, and activities all have a research phase, a selection boundary, an external commitment, and a post-action confirmation. The fields differ, but the operating pattern remains stable: capture evidence, expose tradeoffs, obtain approval where required, and verify the resulting artifact.
Next steps: make one booking path reviewable
Start with one travel workflow and define its state transitions before adding more providers. Save a small evidence record for every serious candidate, separate hard constraints from preferences, and require a pre-booking summary that a traveler can approve in plain language.
Then test the edges: a changed price, an expired session, contradictory baggage information, and a timeout after submission. The goal is not to make the agent sound certain. It is to ensure that every recommendation has a traceable basis, every external commitment has a clear gate, and every uncertain outcome has a safe reconciliation path.
That is the companion lesson to a drainable travel agent: research can be broad, but booking must be narrow, evidenced, and explicit.