I sent more than fifty agents at a website and it came back looking like dogshit.
That is not a figure of speech, it is roughly what I wrote at the time. Monday 23 June, on the Sales Compass site. I had asked for what sounded like a sensible piece of work: convert the pages to proper Gutenberg blocks, strip out the inline styles, improve the semantic structure. Good hygiene. The kind of thing you postpone for a year.
So I fanned out agents. Five per page. One for aesthetics, one for code, one to verify. Then more. Then another dozen or so on top of that.
What came back was technically closer to correct and looked nothing like the site I had built. Here is what I actually typed:
"my point is the site looks like shit and we may have to rollback… you had over 50 agents fan out, with another dozen or so and it looks like dogshit, it should be monumentally awesome for the amount of carbon you released into the atmosphere"
I stand by the carbon line.
The temptation is to conclude that the agents were bad, or that the model was bad, or that this whole way of working does not hold up. None of that is what happened.
What happened is that I gave fifty workers a brief that would have been ambiguous for one, and then had no mechanism for noticing that they were drifting until all of them had finished drifting in slightly different directions.
Nobody was wrong. There was no shared definition of right.
This is the failure mode nobody writes about, because it does not look like failure while it is happening. There are no errors. Nothing crashes. The tests, such as they are, pass. Every individual change is defensible. You only find out at the end, when you look at the thing as a whole and it is not yours any more.
If it had been one agent I would have caught it in ten minutes. Fifty agents produced fifty small reasonable decisions that added up to something unrecognisable, and by the time I saw it the only sensible options were roll back or spend days unpicking it.
Two things came out of that day, and they are the spine of everything that follows.
The first is that scale without structure is not leverage, it is just a faster way to get somewhere you did not want to go. More agents on a vague brief is worse than one agent on a clear one.
The second is that I had no artefact anywhere that said what the site was supposed to look like. It was in my head. Fifty agents cannot read my head. They can read a ticket.
I now cap the fan-out unless the work is genuinely parallel and genuinely well specified. On 8 July I wrote "bumped you up to fable 5 but only one agent ok, no fan out", and that is now closer to my default than the alternative.
Have you had this one? The failure that produces no errors at all, and that you only see when you step back and look at the whole thing?