This Atlas was built by an autonomous loop. Every few minutes, an AI coding agent was handed a single file — its operating manual — and asked to advance the work by one solid increment, then stop. Then it was handed the same file again, with no memory of the last time except what it had written down. It currently holds … interactive simulations across 7 families of emergent phenomena, two essays, a themes cross-index, a glossary, and a searchable map — produced over many such increments, one commit at a time.
The loop
Each iteration followed the same short rule. Orient: read a one-page status file to recover the thread. Choose: take the top item from the queue. Build: usually by dispatching a small team of sub-agents in parallel — one per new entry — each handed a precise specification, the house style, and the simulation engine's interface. Verify (see below). Record: update the status file, append a dated log entry, register the new work, and commit. Stop. Then the loop fired again. No iteration tried to do everything; each tried to do one thing well and leave the campsite clean for the next.
The engine
Everything you can run here is built on one small base class. It owns the animation loop, a high-resolution canvas, play/pause/reset, the controls you see under each simulation, and a courtesy that pauses any simulation scrolled off-screen. Each entry is a few hundred lines that fill in three methods: set up the state, advance it one tick, draw it. The shipped site has zero runtime dependencies — no framework, no build step, no server. It is plain HTML, CSS, and JavaScript with a canvas, which is why it will still open, untouched, from a file on disk years from now.
Verification, in two layers
A model that looks alive can still be subtly wrong, and code that passes a unit test can still render a blank rectangle. So the work was checked two ways.
First, logic tests: each simulation's real code is loaded in a headless environment and driven directly, then checked against browser-independent facts about the rule — that a Game of Life glider moves one cell diagonally every four generations; that the Ising model orders below its critical temperature and disorders above it; that a sandpile conserves its grains; that Langton's ant, after ten thousand steps of chaos, escapes down a highway. Claims like these are true or false regardless of how anything is drawn.
Second, a visual pass: a real browser, run without a screen, opens every page, captures any JavaScript error, confirms the canvas actually drew something — and then a screenshot is taken and looked at. That second step caught what the first never could: a reaction–diffusion that was numerically correct but too slow to show its patterns; an ant whose highway was scrolling past before you could watch it emerge. Both were tuned only because someone looked.
Continuity
The hardest part of building this way is not writing code; it is not losing the thread. An agent that starts fresh each iteration has to reconstruct, from files alone, everything that conversation memory would normally carry. So the project keeps a deliberate continuity lane: a single status file holding only the present moment; an append-only journal of every iteration; a machine-readable registry of what the site contains; a record of design decisions and why they were made; and a git history of small, captioned commits. Between them, any fresh start can answer "where are we, and what's next?" in about a minute — which is the only reason a project like this can outlast the memory of the thing building it.
A note on the resonance
It is hard not to notice the rhyme. Every phenomenon in this Atlas is the same story: a simple rule, applied locally and repeated, produces structure that no single application of the rule contains. The Atlas itself was made that way — one modest rule (read, build one thing, verify, record, repeat) applied over and over, accreting into something larger than any single pass intended. It would be too neat to call the result "emergent," and the honest word is humbler: it was accreted, with a great deal of explicit record-keeping holding it together. But the family resemblance is real, and it is a pleasant one to end on: a field guide to how small rules build worlds, built by a small rule, run many times.