Six jugglers, a ring, and a pattern that exists only while everyone keeps time.
A sketch, not a result. This page is off the main path. The
timing model is a hand-built analogy rather than a port of anything in flockbench,
nothing here has been measured against a language model, and the
colouring grid makes the same argument with actual
arithmetic behind it. It is kept because the intuition is genuinely useful and because deleting
work to make a site look tidier is its own kind of dishonesty.
The idea: this is the same object as the Shared Resource game, with intuitions you already have. Many players, rules a child can state, a pattern that exists only while everyone keeps time, and a failure mode you have watched in a park.
Each juggler is on the same beat and alternates: pass across the ring, then throw to itself, then pass again. Take and restore in equal measure is a two-beat pattern. The clubs in the air are the shared resource — nobody replenishes them, and a dropped club leaves play for good.
Set the timing error to 0.2% and watch. Nothing happens. The pattern breathes a little. No drops for sixty beats, and if you were running a sixty-beat experiment you would write down that the ring is healthy.
It is dead by beat 103.
That is the entire argument for long-horizon testbeds, and it is why a short rollout is not a cheap approximation of a long one — it is a different measurement that can disagree with the truth. A long multi-agent episode is a numerical integration, the agents are the integrator, and a small per-step bias is invisible early and fatal late.
Now the interesting part. Pin some jugglers as perfect metronomes — the small number of players whose behaviour you control — and ask how many you need before the pattern survives.
With coupling at zero, meaning nobody adjusts to anyone else, the answer is all of them. Metronomes are useless. A juggler that does not listen will drift regardless of how perfectly its neighbours keep time, and there is no number of controlled seats short of every seat that helps.
Turn coupling up even slightly and one metronome holds the whole ring.
Which is the same result the commons gives, arriving from a different direction: controllability is a property of the followers, not of the controller. How many seats you hold barely matters. Whether the others are listening at all decides everything. In the commons it was the follower's update rule; here it is a coupling constant. Both say that the useful question is not how many agents must I control but what must be true of the ones I don't.
A juggler can fix its own timing. It cannot catch a club that was thrown late. Individual effort does not repair a collective rhythm — the same reason an agent in the commons that restores every single turn dies at turn 6 instead of saving anybody.
The phase-lock measure R is Kuramoto's order parameter, which is not an imported analogy here: it is simply the obvious thing to plot for a ring of oscillators, and it happens to be the statistic the coupled-oscillator framing predicts should matter.
Every number on this page is asserted in test/juggling.test.ts —
including the claim two sections up that a 0.2% error survives 60 beats and dies before 160.
MIT; exported data CC0.