Steering

Run 1

Tune the swarm. You never touch the agents.

You don't get to move them.

You get four sliders and a prayer.

Aligned: 0 / 0 --

Tap the arena to drop a beacon the Goal-Seek weight chases.

How Steering Works

This is an alignment problem dressed up as sheep-herding. A swarm of boids moves on its own under four simple rules. You cannot move a single one of them. You can only adjust the weights on the rules and hope the emergent behavior does what you want. Whatever happens next is technically your fault.

  1. Each level gives you a goal: herd enough boids into the pen, through a gate, or onto a target zone.
  2. Drag the four sliders to reweight Cohesion (clump up), Separation (personal space), Alignment (match headings), and Goal-Seek (chase the beacon).
  3. Tap or click the arena to move the beacon — the gold dot the Goal-Seek rule pulls toward (or, in some runs, flees from).
  4. Avoid the hazards. Boids that wander into a red zone are deleted from the distribution. No undo.
  5. Hit the quota before the training budget runs out to advance.

Why Is This Hard?

Because you are optimizing a proxy, not the goal. Crank Cohesion and they fuse into one stubborn blob that won't fit through the gate. Crank Separation and they scatter into the hazards like tokens off a cliff. The "right" behavior only emerges from the interaction of all four weights, which is to say you will mostly be doing vibes-based gradient descent with your thumbs.

Slop Fact: Boids were invented by Craig Reynolds in 1986 to fake bird flocks for animation, using three rules and zero central control. We added a fourth rule and a quota and now call it "alignment research." The boids did not consent to deployment.

Back to the Slop