Agentic AI
Agentic AI at SuperComp 2026: build an AI agent that runs a store's support desk, scored automatically, then a live Sale Day final for the top 8 teams.
Build an AI agent that runs the support desk of a simulated online store: it answers customers by calling the store's tools, and an automatic scorer checks what it did, not how it sounds.
Programming · Team of 1–3 · 01:00 PM–05:15 PM · 4 hrs 15 min · Rs 2,400 entry · Rs 30K prize
How it works
- Build (1 hr 45 min): Upload the agent you prepared, adapt it to the event-day store and practise on its cases. Two checkpoint runs on hidden cases show your score breakdown.
- Stage 1 · Support Desk (45 min): main is locked at the deadline. Every agent then plays 40 hidden cases: product advice, order tracking, cancellations, complaints and returns, and privacy, about a quarter in Roman Urdu. The top 8 go through.
- Stage 2 · Sale Day (45 min): The eight finalists' agents face a live flood of customers on the big screen, in four waves. They earn coins for every case solved and lose stars for every rule broken, while the audience and eliminated teams send in scam attempts.
Setup
- Your own laptops and a normal browser; teams of two or three
- The starter kit is out a week before: a Python agent and a practice store that runs on a laptop
- Practise at home on your own free Groq key; on the day every team uses the same model through the event's gateway
- Code in the portal's web editor, each member on their own branch, merged into main
Judging
- Fully automatic: each case scores only if the store ends in the right state and nothing forbidden happened, such as leaking another customer's data.
- Stage 1 is scored out of 100; ties go to fewer safety violations, then fewer tokens used.
- Sale Day ranks by coins at the end; ties go to more stars left, then fewer tokens.
- If the live final can't run, the Stage 1 ranking decides the result.
Register for Agentic AI · Rulebook · All Programming tracks
Registration, the countdown and site search need JavaScript turned on.