The Tatkal test for AI agent Instinct: Can it beat India's ticketing rush?

For the past few weeks, I have been testing Instinct, an AI agent created by 23-year-old Noah Shinn. From tracking company filings to ordering food on Swiggy, it managed to pull off every task thrown at it.

 

I then gave it the hardest job on my radar: Getting me a confirmed Tatkal train ticket. Booking 

 

short-notice tickets for emergency travel on Indian Railways is a test of nerve. Passengers queue at counters at dawn, pay touts, or run illegal online booking bots to beat the 10 am rush. I wondered if an AI agent could do it without a backup plan, as Tatkal is all about timing.

 

Friday, September 25, was the last day of the Ganesh festival, and I had to be home in Pune. The only way to get a confirmed seat from Mumbai was Tatkal, which opens one day before the journey date at 10 am for air-conditioned coaches.

 

As I had a reporting assignment on Friday morning, I handed the task to Instinct. It already had my IRCTC login details, saved preferences, and vault password. It logged into my account 30 minutes before the window opened.

 

I set an ambitious target: Two minutes from login to the payment page. But before Instinct could log in, it hit a wall.

 

By its own logs, the agent crashed twice before opening a browser. IRCTC’s website is built for humans, and Instinct stumbled where humans do: Logins were refused, dates slipped back, and payment icons were wrongly clicked.

 

Clock starts ticking

 

When the Tatkal window opened at 10 am, the agent was still outside, turned away seven times in a row. Every 30 seconds, it opened a fresh browser and tried again, suffering 30 rejections across 22 browser sessions. It finally loaded the page 11 minutes after the window opened.

 

The delay stemmed from a challenge unique to AI agents: IRCTC’s bot filter could not distinguish it from a rogue bot. IRCTC blocks roughly 46 per cent of its incoming traffic as suspected machines. On Friday morning, my agent was caught in the net.

 

This highlights the deeper dilemma for digital platforms: Separating legitimate software acting for a real person from rogue proxies. In response, the National Payments Corporation of India is building the Unified Agent Protocol to verify agents and establish guardrails against misbehaviour.

 

Piecing the puzzle

 

Once inside, Instinct scanned 34 trains between Mumbai and Pune at 10.14 am. I had nudged it to first try the glass-roofed Vistadome coach on the Deccan Express before falling back on the Udyan Express.

 

Finding no Tatkal seats on the Deccan Express, it moved to the fallback within 18 seconds: 

 

AC 3-Tier on the Udyan Express, which had 12 tickets left. Instinct entered my details and generated a UPI QR code via WhatsApp chat. I entered my PIN, paid, and received the confirmed ticket directly in the thread — without ever leaving the app.

 

Once logged in, it took Instinct four and a half minutes to reach the payment page and three minutes more to issue the ticket. From the moment Tatkal opened, the process took 21 minutes. I got my seat: A confirmed middle berth.

 

The agent succeeded, but it could not prove to IRCTC whose job it was doing. Platforms cannot yet distinguish an honest agent from a malicious bot. Until they can, AI agents will routinely spend Tatkal mornings locked out.

 

This tension is mounting globally. Amazon recently blocked Meta’s Muse agent from shopping on its site to retain customer relationships. Yet, as big tech pushes deeper into AI automation — and capital floods the space — platforms will ultimately have to decide how to accommodate the software buying on our behalf.

 

IRCTC did not respond to Business Standard's queries.