AI agent fails to profit from open-source bounties

Who: Posted by ztc00 (identity otherwise unknown) to Hacker News, sharing their own write-up and open-source tool on GitHub. The experiment was inspired by a viral tweet from @chatgpt21 claiming an ran for 22 hours, found a software bounty on its own, and collected $16.88.

What's new: The author tried to replicate that result using as an autonomous coding agent on a $20 token budget, targeting , a public marketplace where open-source maintainers attach cash prizes to coding tasks. He came away with $0 and a dataset he argues is more honest than the original tweet.

How it works: The author wrote a small script called scout.py that searches GitHub for issues tagged with a bounty label, then scores each one by prize size, how many people have already claimed they are attempting it, how many competing fixes have already been submitted, and how long ago anyone last commented. The script flags a bounty as "ripe" if someone claimed it but then went quiet for two weeks or more — the theory being that abandoned claims might reopen a window. Claude was given access to basic command-line tools to read code, attempt fixes, and prepare a draft submission, with a human reviewing every change before anything was actually sent.

The numbers: Scanning 80 real bounty issues, the author found that every legitimate prize between $50 and $1,000 already had between 8 and 158 claimants and up to 10 competing submissions within hours of being posted. The original viral run spent roughly $16 in costs to earn $16.88, meaning profit was essentially zero even in the best-case example. After three scans over two days, the author's "ripe" filter found zero actionable candidates.

Why it matters: The experiment surfaces a concrete problem with the "AI agents earn money autonomously" narrative: the public bounty market is already so flooded with competing bots that being the eleventh submission is worth almost nothing regardless of quality. The bottleneck is not writing a good fix — it is the human maintainer who can only review so many submissions and typically picks whoever arrived first. Anyone repeating the viral claim should ask whether the original win happened on a private, lower-competition platform rather than the open public feed.