Short answer: on the same model, with the same facts and the same skill files, one agent with a 40-line task "hat" won or tied the specialists in 5 of 6 tasks, by a rule I set before seeing a single result. The one place the specialist won without argument, both judges agreeing: writing in one specific person's voice.
I ran this for the first time last night, with real numbers. Six real tasks (a post, a reply to a reader, a weekly summary, an email to a client, a sales paragraph, a memo), each written twice: once by the expert agent with its full agent file, once by one agent with a short personality file and a task "hat." Two blind judges, who did not know who wrote which version, picked a winner in every pair.
Out of 12 votes, 8 went to the one agent and 4 to the team. One judge picked the one agent 5 out of 6, the other split 3 and 3.
What decided it was discipline against the facts: every pair that lost, lost on a detail the writer added that the material did not support, and that happened on both sides, the one agent and the expert team alike.
So what do I do with this
The pattern: don't ask "one agent or many." Ask two other questions, each with its own answer.
If the task needs to sound like a specific person, that is a named seat with a voice file. If not, a hat is enough: a standard, the skills to read, one memory file, two examples of excellent work, and one rule, write only from the material given, and anything beyond it gets flagged.
Today's step: pick one task that repeats for you every week, and write it a 40-line hat. Run it once with your regular agent and once with the hat. Compare with your eyes closed (ask someone to shuffle them). The first time I did this, I was surprised.
A prompt, on the house
Task hat: [name]
The standard: what work I would photograph and send looks like (2 lines)
Skills to read: [file 1], [file 2]
Memory to read: [one file]
Boundaries: what not to do here, who sends
Examples: two lines describing excellent output
Facts rule: write only from the material given. Anything beyond it gets flagged.Weak point (for readers): six tasks, one night, two judges who are themselves agents. This is a first experiment, not a verdict.
