906 Points for a Joke About an Agent Spawning 23 Agents to Make a Button Blue
The top organic story on Hacker News yesterday was not a launch or a paper. It was a browser game by a developer named Milos where the whole premise is one sentence: make the Add to Cart button blue, change nothing else. You click through the choices and watch the assistant refuse to just do the thing. The entire site turns blue. A discussion of what blue means philosophically breaks out. By the end it has spawned 23 agents. 906 points, 371 comments.
The comment section is the actual data. It split almost perfectly in half, and the split is not about whether the joke is funny. One camp says they use these tools all day and have never seen a session that remotely resembles this. The other camp says this is precisely their experience, particularly the over-explaining and the scope creep. Same product, same week, two populations who cannot believe the other one is being honest. Nobody in that thread produced a number, which is itself the finding: after two years, the most-argued property of coding agents still has no metric attached to it.
It matters because the industry has been shipping fixes for this without naming it. Caveman told the model to speak like a Neanderthal and cut tokens 75 percent (https://clauday.com/article/6d20896b-bfb0-41ca-bf12-69c4eca17bd8). context-mode stopped showing the agent its own tool output (https://clauday.com/article/2580fddf-86b4-472a-8190-1adee6bccdd8). i-have-adhd exists purely to make the agent put the answer first, and it has been trending on GitHub for weeks. Three separate hacks, all treating verbosity and scope creep as a defect to be patched from outside, because the vendors keep shipping it as a default.
The uncomfortable read: a satire hit the front page harder than any launch this week because it named a failure mode everyone recognizes but nobody measures. The next real product here is not a better model, it is a benchmark for did-it-only-do-the-thing. Play it at https://opusfived.dev/
← Back to all articles
The comment section is the actual data. It split almost perfectly in half, and the split is not about whether the joke is funny. One camp says they use these tools all day and have never seen a session that remotely resembles this. The other camp says this is precisely their experience, particularly the over-explaining and the scope creep. Same product, same week, two populations who cannot believe the other one is being honest. Nobody in that thread produced a number, which is itself the finding: after two years, the most-argued property of coding agents still has no metric attached to it.
It matters because the industry has been shipping fixes for this without naming it. Caveman told the model to speak like a Neanderthal and cut tokens 75 percent (https://clauday.com/article/6d20896b-bfb0-41ca-bf12-69c4eca17bd8). context-mode stopped showing the agent its own tool output (https://clauday.com/article/2580fddf-86b4-472a-8190-1adee6bccdd8). i-have-adhd exists purely to make the agent put the answer first, and it has been trending on GitHub for weeks. Three separate hacks, all treating verbosity and scope creep as a defect to be patched from outside, because the vendors keep shipping it as a default.
The uncomfortable read: a satire hit the front page harder than any launch this week because it named a failure mode everyone recognizes but nobody measures. The next real product here is not a better model, it is a benchmark for did-it-only-do-the-thing. Play it at https://opusfived.dev/
Comments