Thirty agents, far fewer than thirty opinions
Adding agents to a team stops adding independent judgement long before it stops adding cost. These are the measurements behind that claim: every open model we tested, on every reasoning benchmark, under debate.
–open models, 3B–27B
–model × benchmark cells
–effective team size at N=30
–cells in near-saturation
Effective team size vs nominal team size
Each line is one model. The dashed diagonal is the independent-voter ideal — where N agents would contribute N independent opinions. The shaded gap is what never materialises. Hover a line, or a model chip, to isolate it.
3B27B
model size — colour
Where each model × benchmark lands
The fitted regime law R(N) = 1 / (1 + c(N−1)N−β).
β near 0 means a hard ceiling: effective size stops growing regardless of how many
agents you add. Colour is the benchmark; hover for the model.