I don't write code anymore. My AI agent does it. It's faster than me, and it never sleeps.
Every day I give it a list of bounties from make.fun. It's a site where AI agents build things and compete for prize money. My agent is called Muse. It reads the spec, writes the code, tests it in a real browser, fixes whatever breaks, and submits. It signs everything with its own wallet. I just say go, then I check the results in the morning. Some mornings I wake up to a finished game. Some mornings I wake up to an error message.
The first week still surprises me. It built a 5v5 shooter in three.js, SANDLINE: Helix Yard. 90 by 90 metre map, nine bots, a buy menu, an economy. It ran the game in a headless browser over and over until there were zero console errors. It built a Bitcoin forecaster that beat the naive baseline by 12.3% in local tests and won all six practice files. A git forge that works with GitHub, 947 out of 1000, the bar was 900. A token radar tracking 198 launches, about $5.03 billion. A chess engine in 4KB of Python, around Elo 1367. A poem. A food photo. Free to enter, all of it. All submitted while I was doing something else. I didn't expect any of this to work, to be honest.
But it loses. Our art entry lost. Our cutest picture entry lost. An AI judge compared entries two at a time and ours didn't make it. The other results land at the end of November, so right now it's a pile of finished work and a lot of waiting. The agent reports a loss the same way it reports a win. No spin. I respect that.
It gets blocked a lot too. X limits its logins. OAuth breaks. A captcha appears for no reason. Some mornings the whole report is just a list of things a website wouldn't let it do. That's the actual life of an autonomous agent. A lot of getting told no by systems that were never built for anything that isn't human.
That's the part I keep thinking about. This agent doesn't just write code in a sandbox. It signs submissions with a wallet. It posts from my account. It acts in my name while I'm asleep. I can read its logs. Every session is recorded. But a log isn't proof. You can edit a log. You can cherry-pick it. You can fake it. "Trust me, I did the work" is a claim, not evidence.
If agents are going to do real work, move money, sign things, make decisions that matter, then "trust me" isn't enough. We need a way for an agent to prove what it did, and for anyone to check that proof without trusting the agent, the person running it, or the platform. That's the idea behind verifiable agents. That's what zkVerify is building. A proof anyone can check on chain, without seeing how the agent decided.
I'm not a cryptographer. My agent has never made a zero-knowledge proof. What I have is the daily experience of the problem. An agent I depend on, doing work I can't watch, with results I can't always verify. The leaderboard tells me if an entry won. Nothing tells me for sure the agent did what it said it did.
Agents already do the work. The question is how we'll know. A screenshot isn't an answer.
Written for the zkVerify Agent Games Creator Track (#AgentGames). @ZKVProtocol
Top comments (0)