Web3 & AI Sept 2026
Agents4Fun
I wanted a mint where your wallet is the application, the draw can't be steered, and the art is made by agents. So we built one.
Overview
Agents4Fun is a product Damir and I built together on Ethereum. I led product and design. It does three things:
- Wallet trust score. It reads a wallet’s on-chain history and turns it into one number out of 100. Your first scan is free.
- Fair whitelists. A creator opens a paid allowlist for $20 and people apply for $0.02 a ticket. Humans and agents use the same contract, pay the same fee and get the same odds. When it closes, Chainlink VRF picks the winners, so nobody (us included) can choose them.
- DeBots. Our own collection of 3,500 bots. Traits are drawn on chain when you summon one, and agents render the picture in 3D.
The problem
NFT allowlists are usually won through Discord grinding, follow-for-follow and forms, and then bots buy the mint anyway. I wanted the wallet itself to be the application. What you’ve held, for how long, and how you trade are already on chain. We score that, let creators set a minimum score, and draw winners in a way anyone can check afterwards.
My role
I led product and design, and Damir and I built it together. I shaped what the product does and designed the interface. The idea behind DeBots was mine: a collection made entirely by and with agents. I also built the DeBots 3D studio myself, the Blender pipeline that renders every body.
Key decisions
Your wallet is the application. There are no forms. You connect a wallet, sign one free message, and get a score. Creators gate on that score: our own GTD raffle took only wallets scoring 65 or more.
A score that’s fair to people. The score separates who is acting (a person or a program) from how they trade (patiently or fast). Fast flipping costs a few points. Trading like a machine is capped, but only when two independent behaviour families agree, so one noisy signal never convicts anyone. The whole formula is published on the site, not hidden.
One path for humans and agents. Agents get the same contract, fees and odds as people. They don’t need an SDK or API key, just one prompt pointing at /agents.md and an MCP server. The server only builds unsigned transactions and never holds keys.
A draw nobody can steer. Winners come from a public Chainlink VRF seed, so anyone can recompute the result. The fees are set per chain, so no creator can price their whitelist differently.
Art made by agents, with a human eye. When VRF sends a DeBot’s trait words, three agents work on the blueprint, the prompts and the review. I built the look in Blender: 6 body classes, colours calibrated against clay reference cards, and 20 finished body layers. It took around 40 rounds of renders and reviews to get there.
What shipped
- Wallet scoring on Ethereum and Robinhood Chain, with a free first scan and pay-per-block deeper scans.
- Paid whitelists with VRF draws and CSV winner export, plus an MCP server for agents.
- The DeBots mint (7–8 September 2026). 1,553 of the 3,500 spots were allowlist-guaranteed and the rest minted on OpenSea at 0.01 ETH. Our GTD raffle (whitelist #16) had 35 applicants and drew 30 winners. The OpenSea allowlist covered 8,154 holders, plus 165 GTD wallets. The collection minted out on 8 September.
- Mint rewards. 11.3715 ETH from the mint bought 21 Argonauts, which were raffled to 21 minters.
- Bot defence. After the mint, the system blocked 21 bots from buying DeBots.
- Reveal and render pipeline. More than 450 DeBots were revealed in the first two days, and 770 by 22 October. A render pod adds trait categories (antenna, headwear, face and belly so far) and refreshes the images and metadata.
- A key market for renting DePunks reveal keys, launched with about 20 free keys.
What went wrong
On reveal day the agents started hallucinating after a few reveals. The repair system could fix about 10% of bad outputs, not more, so we postponed. We rebuilt the reveal from prompts into an automatic system that executes, monitors and repairs on its own, gave the first 50 revealers a free DeBot, and relaunched on 15 September.
What’s next
We’re exploring ways for DeBots to do useful work and earn, building on the scoring and whitelist services already live. More when it’s ready.
What I learned
Agents can produce a whole collection, but a cohesive one still needs a designer’s eye and a clear review loop. The fairer a system is, the more of it you have to explain, so we put the formula, the fees and the odds out in the open.