∗ Guide

Inside an agent task: how TowerAI books a train ticket

It's easy to describe an AI agent in a sentence. It's more useful to watch one work. In this guide we follow a single, ordinary request through TowerAI's booking agent, from the moment you ask to the moment the booking is confirmed, one stage at a time. Along the way we'll show where the agent stops to ask for your approval, and what it writes in the activity log.

The example is illustrative. Trains, timings and screens are simplified to show the shape of a task rather than the details of any particular service, and real rail sites will look different.

The request

You're planning a Saturday in Mysuru with a friend. You open TowerAI on your phone and say:

"Book two train tickets from Bengaluru to Mysuru this Saturday morning. Window seats if possible, and we need to be there by 11."

That's one sentence. It carries a route, a date, a number of travellers, a deadline and a preference, and it leaves a few things unsaid. Here's how the agent gets from there to a confirmed booking.

Stage 1 — Ask

Ask is simply the request, in whatever form is easiest. You can say it, type it or show it; a screenshot of a timetable or an event invitation works as well as a sentence. Here it's spoken, so speech-to-text turns your voice into text first. Nothing has been searched or booked yet. The agent has only listened.

Stage 2 — Understand

Next, the language models work out what you mean. A casual request is full of implied detail, and the agent's job is to make it explicit:

  • Route — Bengaluru to Mysuru, one way.
  • Date — "this Saturday", resolved to an actual calendar date.
  • Travellers — two.
  • Hard limit — arrive in Mysuru by 11:00.
  • Preference — window seats, ideally together. "If possible" makes this a preference, not a deal-breaker.

Where something is genuinely unclear, the agent asks instead of guessing. If it doesn't have the details it needs for the second traveller, it asks for them now rather than discovering the gap halfway through a booking form. A quick question at this stage is far cheaper than a wrong ticket later.

Stage 3 — Plan

With the intent pinned down, the goal becomes a list of concrete steps. For this task, the plan might read:

  • Search rail booking sites for Saturday-morning trains on the route.
  • Drop anything that arrives after 11:00.
  • Compare what's left on departure time, journey time, class and seat availability.
  • Choose the best match for two travellers with window seats, and keep a fallback.
  • Fill in the traveller details, then stop at payment and ask for approval.

Note the last step. The pause before payment isn't bolted on at the end; it's written into the plan from the start. And because the plan is visible to you, you can redirect it early. If you'd rather take the earliest train regardless of seats, now is the time to say so.

Stage 4 — Act

This is the legwork you'd otherwise do yourself. The agent opens rail booking sites, enters the route and date, reads the results and applies your constraints. Suppose the search turns up three morning trains (all details illustrative):

  • Train A — departs 06:30, arrives about 08:45. Window seats available for both travellers.
  • Train B — departs 08:00, arrives about 10:30. Only one window seat left.
  • Train C — departs 09:15, arrives after 11:00. Set aside, because it breaks the hard limit.

Train A meets every part of the request, so the agent selects it and holds Train B as the fallback. It fills in the traveller details, picks two window seats and works through the site's booking steps until it reaches the payment page.

Then it stops.

Stage 5 — Check in

Check in is the stage the whole loop is built around. Before anything that spends money or speaks for you, the agent asks. On your phone you'd see a short summary, something like this:

  • Ready to book: 2 tickets, Bengaluru → Mysuru, this Saturday.
  • Train: Train A, departs 06:30, arrives about 08:45.
  • Seats: two window seats.
  • Total: the full amount shown on the booking page, with any fees listed separately.
  • Your options: approve, switch to Train B, or cancel.

Nothing is paid until you approve. When you do, the agent completes the payment and waits for the confirmation. Then it reports back with what you need for the journey: the booking reference, the train, the timings and your seat numbers.

Two details are worth knowing. First, a check-in can happen mid-task, not only at the end; Act and Check in take turns whenever the agent reaches a step that needs you. Second, if something changes while the agent is waiting — seats on popular trains can go quickly — it tells you and offers the fallback. It doesn't quietly book something you didn't approve.

What lands in the activity log

Every step is recorded in the activity log, so you can see what the agent did, in what order and why. For this task, a simplified version of the log would look like this:

  • Ask · Request received by voice: 2 tickets, Bengaluru → Mysuru, Saturday morning, arrive by 11:00, window seats preferred.
  • Understand · Date resolved. Asked for the second traveller's details; received.
  • Plan · Five-step plan created, with a pause before payment.
  • Act · Searched rail booking sites. Three options found; one set aside for arriving after 11:00.
  • Act · Selected Train A (06:30), kept Train B as fallback. Entered traveller details and seat preference.
  • Check in · Paused at payment. Approval requested.
  • Check in · Approved by you. Payment completed; booking confirmed.
  • Check in · Summary sent with booking reference and seat numbers.

The log is handy when everything goes right. It matters more when something doesn't. If a site fails halfway through, or a seat disappears between search and payment, you can see where it happened and exactly what the agent did next.

The agent can take every step of the booking for you. It just won't take the one that spends your money without asking first.

When things don't go to plan

Real websites are messy, and AI agents can make mistakes: misreading a page, say, or picking an option you wouldn't have picked. That's why the check-in comes before payment rather than after it, and why every step is logged. It's also why you can pause or stop an agent at any time. If you've changed your mind about the trip, or you'd simply rather finish the booking yourself, you can take over.

The same loop, other tasks

Change the request and the loop stays the same. A call agent rescheduling an appointment, a personal assistant drafting a message and a workflow agent running a multi-step process all move through Ask, Understand, Plan, Act and Check in. In every case, the check-in comes before anything that spends money or speaks for you. There's more on each agent on the agents page.

A short checklist for good results

  • Say what matters most: time, route, seats or price.
  • Separate limits from preferences. "Arrive by 11" is a limit; "window seats if possible" is a preference.
  • Answer clarifying questions early; they save rework later.
  • Glance at the plan before the agent starts acting.
  • Read the check-in summary properly before you approve.
  • Use the activity log to see what happened, especially if anything looks off.

Booking a train is a small task, but it shows how every TowerAI agent works. If you'd like to try one yourself, start for free — the Personal plan is free — or talk to us about agents for your team.

Try an agent task of your own.

Start with something small: a booking, a reminder, a call. TowerAI plans the steps, does the legwork and checks in before anything that spends money or speaks for you.