In the first post I said the bread and butter of all this was sending an agent out on a research mission. So here is the obvious next question: if one mission is good, why not run ten at once?

The early version was not impressive. One agent, one question, a list of search results, and a lot of patience. It would search, read, search again, and eventually hand me something. It worked, but it was slow: a single agent reading one source at a time tends to fall for the first decent answer it finds and stop looking.

So by April 2026 I built the thing you are picturing. Instead of one agent chewing through a question end to end, I broke the question into pieces and handed each piece to its own agent, all running at once. One on the market, one on the technical details, one on what the critics say. A final pass stitches their findings into a single answer.

For the right kind of question, this works well. Faster, wider, more thorough. It would be easy to stop there and decide more agents is always better.

It is not. I learned the hard way that fanning out only helps when the pieces are independent.

Some questions are a wide field: go check these eight unrelated things, none depending on the others, and bring it all back. A crowd handles that easily. Other questions are a chain: what you find in step one changes the question you ask in step two, which changes step three. The whole thing is one train of thought.

Split a chain like that across a crowd and it gets worse, not better. Each agent sees only its slice, and nobody can follow the thread running through the whole problem. You get five fragments that do not add up to one line of reasoning. It is like giving five people one chapter each of a mystery novel and never telling them the ending. Each chapter reads fine on its own, but the book makes no sense.

So the rule I settled on is simple, and it cuts against the marketing. Parallelism is a property of the question, not a virtue of your setup. If the parts are independent, fan them out. If each step feeds the next, keep it one agent thinking hard for a long while, and do not reach for the fleet just because you have one.

There is a second thing nobody warns you about. When you fan out, the hard part moves. It stops being the searching and becomes the putting back together. Ten agents handing you ten piles of findings is not an answer. It is ten piles. The step that turns those piles into one clear, defensible conclusion is where the real work, and most of the mistakes, now live. I used to think gathering was the job. Gathering is the easy part.

Early on I caught myself about to throw a whole fleet at a question, mostly because I had one and wanted to use it. The question was a chain, not a field, and I already half knew the answer from earlier work. One agent, one afternoon, and a careful look at what we already knew beat the crowd I almost launched. Knowing when not to use the heavy machinery turned out to matter as much as having it.

Which raises the question I kept avoiding. If the answer is so often already sitting in work we did weeks ago, why are we researching it again at all? That is the next post: the day the library stopped being a graveyard and started paying us back.