Mike Hale Superteam Korea has published its recap of the Google Cloud and Solana AI Agentic Hackathon Demo Day, held August 21 at Google Startup Campus in Seoul, where 10 finalist teams out of 160 builders competed for a $5,000 grant pool from the Solana Foundation.
Every team was set the same question. Can an agent judge, pay, and settle onchain without a human in the loop?
Google Cloud and the Solana Foundation co-hosted, Ewha Chain organized, and Four Pillars served as research partner. Judges came from nine institutions, including Shinhan Card, KG Inicis, KG Financial, and Stage Five alongside the hosts. Korean card networks and payment processors sent people to score agent settlement demos.
The three winners
The $5,000 was split across the top three, and half of each award is contingent on the team entering the Crypto World's Fair Hackathon later this year.
Place | Project | What the agents do | How it settles |
|---|---|---|---|
1st | Solply (lead Kim Tae-woong) | Franchise branch and HQ agents negotiate ingredient purchases directly | 99.3% of settlement transactions automated |
2nd | A2AHouse (team WAIaaS) | Buyer-side agent marketplace for purchasing on a user's behalf | Self-hosted delegated wallets that execute only inside preset authorization limits |
3rd | KNOT (team Knot) | Brand and creator agents negotiate campaign terms | Escrow releases automatically once a deliverable is verified |
A2AHouse came from Kang Bo-hyeon, Yu Min-ho, and Jung Chan-hong, and KNOT from Park Min-sung, Choi Ye-won, and Yoo Hyo-chang.
The finalists converged on one design
Ten teams working independently arrived at the same architecture, which the recap summarizes as "judgment is AI, limits are code." Gemini decides what to do, and a program enforces the boundary it has to decide within.
Four patterns showed up across the finalist projects:
Agent-to-agent negotiation rather than one agent acting alone. Both sides of a transaction run their own agent and the two settle terms between them.
Escrow first, with conditional release. Funds move into escrow at agreement and only release when a condition is verified.
x402 micro-transactions for the payments agents make to each other.
Gasless, passkey-based onboarding with no seed phrase in the user flow.
Nobody handed these teams a reference architecture and they landed on the same separation anyway, which suggests it falls out of the constraints rather than out of fashion. For anyone building agent payments on Solana, that is the more reusable output of the event than the leaderboard.
The winners picked unglamorous problems
All three winning teams aimed at B2B inefficiencies that already cost somebody money. Solply went after franchise supply settlement, a category the recap sizes at roughly 26 trillion won a year in Korea. KNOT went after creator payment delays. The recap also lists receivables financing among the problems the field took on.
None of them built a demo that only makes sense if agentic commerce arrives first. They took a settlement process that is slow and manual today, and replaced the manual part. The compliance angle came along with it, since an escrow release tied to a verified deliverable produces an audit trail that a spreadsheet and an email thread do not.
Card networks judged the demos and the teams aimed at Korean B2B settlement, which makes this a more grounded showing than most agent hackathons manage. Whether any of the three ship past the demo is the open question, and the grant structure is built to push them. Half the prize money only lands if they enter the global hackathon and keep building.