# Animating our docs with a swarm of agents

Canonical: https://tiptreesystems.com/blog/animating-our-docs-with-a-swarm-of-agents

2026-06-16 · Martin Weiss

![Video: a swarm of agents pushing cubes into place](/blog-images/animating-our-docs-with-a-swarm-of-agents.mp4 "controls")

I was designing Tiptree Systems documentation and decided it needed animations. For the coding agents docs, I thought a little swarm of agents pushing cubes into place would feel good. I started by just asking codex to animate the motion, but it looked dumb (see: my real and silly codex messages below).

\> “it feels like they are all moving to the right on rails. As opposed to some amount of dispersed shared volition.”  
\> “I want you to read about how fish and birds swarm. That’s what I’m looking for here.”

I know a lot of research on swarm algorithms exists, so I made visual studies and decided on a nice mix of algorithms with nice motion feel. But when I tried to get the swarm to like... actually do work and move boxes.

\> “I’m going to try training an RL model so these guys work together to [move the cubes]. I don’t have high hopes for that direction though.”

Training a multi-agent system is hard, and I figured it wouldn’t run on a laptop anyway. But that was 2018 Mujoco spider RL thinking, not 2026 Codex thinking.

So we wrote a 2.5D contact-physics simulator for some cubes that move only when agents touch them. Then trained a tiny shared policy MLP using evolutionary methods.

Evolutionary methods need a ton of rollouts, so wall-clock time on the sim imposes a ceiling on the agent performance. I ended up rewriting my physics sim as a fixed-array C++ evaluator and got it to `12,288` full 1100-step rollouts in `63s`. Once it was that fast, I could give vague design notes to codex (“less drift,” “stop running away from the cubes,” “less jitter at the end”) and it would implement reward shaping and change the agent observation space and retrain the little guys.

The final policy is an 81-input, 8-hidden-unit MLP. It solves a 3x3 cube grid, then reforms the swarm afterward.

Just wanted to share this little view into my personal singularity with y’all.

*First posted on [LinkedIn](https://www.linkedin.com/feed/update/urn:li:activity:7472634173418299392), June 16, 2026.*
