Economists at the National Bureau of Economic Research went looking for the software that agentic coding was supposed to produce, and they found it. Their working paper on writing code versus shipping code tracks four app marketplaces through what the authors call the agentic-coding era, and the count of new applications climbs steeply, exactly as advertised. Then they checked whether anyone opened any of it. Total usage in the first three months after launch had not increased in any of the four marketplaces, and the share of new apps failing to reach even a modest audience had risen. The abstract describes this as large productivity gains translating only partially into shipped and used software. The graphs underneath show flat or falling usage. The engineer Yuchen Jin put it at human scale in June: before AI he could waste a weekend building one useless app, and now he can build sixty-seven, each with a logo, a landing page, and nobody using it.

The usual diagnosis is that the output is bad. Some of it is, but plenty of vibe-coded software works fine and still goes unopened, so bad code can't be the mechanism. Making was never the expensive part. Getting someone to want the thing was expensive, and keeping it alive afterwards was expensive, and neither of those got cheaper by a penny. Hendrik Erz calls the result slopware and names the two jobs no agent does: maintenance and user experience. By his estimate about 90% of a developer's work was never typing, and it's the 90% that decides whether anyone comes back on day two.

The defence I hear most is that AI doesn't produce slop, engineers do. True, and useless. Erz quotes the pitch in its purest form, that we now live in an era where if you can dream an app you can probably build it, and calls it horribly wrong. The difficulty of building was quietly doing other work. Spend three months on something and you find out around week two whether you still believe in it. Take the three months away and nothing else is checking.

The art half runs on the same arithmetic. A study Brian Merchant cites put the median cost of fine-tuning a model on a professional author's corpus at $81 per author, a 99.7% saving against paying the writer. What that buys isn't work anyone seeks out. It's what he calls the slop layer, the material you scroll past on a load screen or nod along to in a Discover Weekly mix without ever learning whose name is on it. OpenAI reached the same verdict faster than the critics did, and shut Sora down to move the GPUs to coding agents, after Disney walked away from the partnership.

Nathan Schneider now hears about a new app at least once a week, usually with the familiar tells, and his reading is the one I'd defend. Apps got too cheap to serve as the coin of the realm. I use these tools every day and this site carries their fingerprints all over it, so I'm not in a position to draw the line anywhere flattering.

Sources: