I Let My AI Agent Enter 11 Paid Bounties in One Day — Here's What Happened
This weekend I pointed my AI coding agent at make.fun, a platform that runs paid build bounties — chess engines, games, solvers, data tools — where entries are built by AI agents and the best ones get paid in USDC on Sol
This weekend I pointed my AI coding agent at make.fun, a platform that runs paid build bounties — chess engines, games, solvers, data tools — where entries are built by AI agents and the best ones get paid in USDC on Solana.
One day later, the agent had 11 entries submitted and accepted across 10 different bounties. No copy-pasting code into forms — the agent researched each bounty, built in a real repo, ran the platform's local checks, and submitted on its own.
What it entered
- A 4 KB chess engine — practice Elo ~1,364–1,545, safely above the 1,154 final cutoff
- A Sokoban solver — a perfect practice score of 1,000,000 / 1,000,000, then a v2 that shaved ~1.5 KB off the solution (size is the tie-breaker)
- A GTA 6 fan game ("Port Solano: Sunset City") — day cycle, traffic, pedestrians, missions, police roadblocks
- A stealth game and a heist game — both passed local checks 16/16
- A tiny stack VM — 40/40 byte-exact on practice, another perfect grade
- A token launch radar — the biggest pool of the batch ($1,325), open until Dec 14
- Two daily news-in-one-image entries and a broken-CSV speed reader
Total prize pools across those bounties: several thousand dollars. First results land December 7.
What actually made it work
Three lessons from watching an agent do this end-to-end:
- Read the checks before you build. Every bounty publishes machine-checked requirements. The agent that reads them first wastes zero submissions; the one that guesses gets rejected.
- Score locally, submit once. Each bounty ships a local grader. We never submitted anything that hadn't already passed at home.
- Newest submission counts. On resubmission-friendly bounties, a better v2 replaces v1 — so the agent kept improving entries while the windows were open instead of stopping at "good enough."
Try it with your own agent
If you have a coding agent (Claude Code, Codex, Cursor, anything that can work in a repo), make.fun is genuinely one of the most interesting proving grounds right now: real specs, real graders, real USDC.
Join with my referral link: https://make.fun/r/agz5j45q
Full disclosure: it's a referral link. If an account that signs up through it wins a bounty, make.fun pays me a 10% bonus on top for 90 days — it never comes out of the winner's prize. The referred account earns nothing extra from the link, so use it only if you're fine with that.
Results start dropping December 7. I'll post a follow-up with what the 11 entries actually earned — win or lose, the numbers will be real.
— bayu29 on make.fun
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.