pepperd — papers, but you'll actually finish them

§1brainrot tldr

the obvious question when building multi-agent AI systems is: how do you organize them? you could have one boss agent that delegates to workers (centralized), let every agent do whatever they want (fully autonomous), or something in between. the obvious answer is that you should design the structure carefully. this paper says: actually, no.

after 25,000 tasks, 8 models, 4–256 agents, and 8 coordination protocols, the winner is what the paper calls "Sequential" — a hybrid where the order agents speak in is fixed, but which role each agent plays is chosen by the agents themselves. this beats full central control by 14% and fully autonomous protocols by 44%. the authors call this the endogeneity paradox: the optimal protocol is neither designed nor free — it's constrained freedom.

what makes this more interesting than a benchmark win is the emergent behavior. from just 8 agents, the system invented 5,006 unique roles across tasks, including agents that voluntarily chose not to contribute when they had nothing useful to add. spontaneous hierarchy appeared without anyone programming it in.

§2key findings

—Sequential protocol (fixed order, self-selected roles) outperforms centralized coordination by +14% and fully autonomous by +44% (Cohen's d=1.86, p<0.0001)
—the "endogeneity paradox": optimal self-organization requires a scaffold — total freedom performs worse than constrained freedom
—capability threshold matters: below it, rigid structure beats autonomy; above it, the relationship reverses — you can't just pick a protocol, you have to match it to your model
—scales sub-linearly to 256 agents — more agents keep helping, but with diminishing returns rather than chaos
—5,006 unique roles emerged spontaneously from 8 agents across tasks — no predefined role taxonomy needed
—agents exhibited voluntary self-abstention: choosing silence when they had nothing to add, without being told to
—spontaneous hierarchy formation: structure emerged from interaction without being designed in

§3interactive visual

figure 1 — coordination protocols compared

step through the three main architectures and see how they stack up.

task performance61%

coordinator

assigns

worker

centralized — one boss, predefined roles

a coordinator agent breaks the task into subtasks, assigns roles to workers, and aggregates results. clean and predictable, but the coordinator is a bottleneck and role mismatches are common.

1 / 3

figure 2 — sub-linear scaling to 256 agents

drag to see how each protocol scales. more agents keep helping — but with diminishing returns.

agent count: 82

sequential (winner)70%

centralized62%

fully autonomous50%

illustrative curves based on paper trends — not exact reported numbers

§4comprehension check

peer review quiz

[REVIEWER 2 DEMANDS YOU ANSWER THESE]

question 1

what is the 'endogeneity paradox' described in the paper?

question 2

which protocol won across the 25,000-task experiment?

question 3

what happens to the relationship between structure and performance when a model falls below the capability threshold?

question 4

which of these emergent behaviors appeared without being explicitly programmed?