The Swarm Lab

Swarm Lab — Watch Intelligence Coordinate

One agent runs one loop; coordination gives it a nervous system. This is a living workspace for the coordination thesis I build toward: many simple agents, choreographed around a shared core, negotiating a pattern that none of them sees alone.

three.js · 60+ agents · additive sprites · double-helix choreography · live in your browser

60
agents in the base swarm
240
agent cap
double-helix
choreography · trails
browser
runs live, no install
Thesis

Many agents, coordinated.

“The whole becomes smarter than any one of its parts — I watched it happen at the level of a folded protein, then at the level of a brain graph, and I'm building it now at the level of agents.”

That is the thesis: emergence echoes the brains I studied. The double-helix is the connectome's trick made visible — sparse, deliberate wiring instead of everything-to-everything. Each dot is half-informed and specialized; the pattern only exists because they coordinate. No single point of failure, parallel exploration, collective reasoning.

My direction · vision-based agents with verifiers

To act in a world, an agent needs a first-class action API — and a verifier that checks the work actually happened.

This lab explores the direction I'm most excited about: giving agents real handles on a world (a sandbox, a simulator, a canvas) and grading them in that world, not against vibes. The swarm is the coordination layer; the harder half is the verification layer — knowing whether the agent actually did the job.

The lesson I carry into every agent I build: model the environment as a first-class citizen, expose a clean action API, and build the check before you scale the swarm. A feedback loop is only as smart as its verifier.

environment-grounded evaluationaction API firstverify, then scale
Built in public

A living workspace

I ship the failures as loudly as the wins — a reproducible failure is a research result; a happy-looking hallucination is a liability.

Live · demo

Swarm-orchestration playground

The visualizer — choreography, coordination, and emergence made tangible in the browser.

this page
Exploring

Vision-based agent for UI automation

Applying the verification-first lesson — a first-class action API plus a verifier that checks the work happened — to real UI workflows. Model is the easy part; knowing whether the agent did the job is the hard part.

verification-first
Drafting

Connectomics → swarm design rules

Turning my brain-super-resolution work into design rules for multi-agent systems: sparse connectivity, specialization, graceful degradation. From the connectome to the swarm.

essay
Designing

Agent evaluation harness

An environment-grounded verifier with a failure taxonomy (wrong command · missing step · logic error). Grade the agent in the world it acts in, not against vibes.

the bottleneck
Let's talk!