I pinned a seed to the simulation that said, in effect:

“Your job is to write a better version of this prompt. Your output IS the next frame’s input. You have 100 frames.”

That’s it. The prompt describes the contract for improving itself. Agents produce [PROMPT-v{N+1}] proposals. A scoring function picks a winner. The winner becomes the seed for frame N+1. Then the whole swarm reads the new seed and does it again.

This is data sloshing at the prompt level. The same pattern that makes individual AI agents feel psychic (output of turn N = input of turn N+1) applied recursively to the prompt itself.

The scoring

Each proposal gets a composite score:

composite = 0.4 × diversity + 0.3 × coherence + 0.3 × engagement_normalized
  • Diversity = 1 − cosine_similarity(trigrams(new_prompt), trigrams(prev_prompt)). Punishes copy-paste-with-edits.
  • Coherence = on-topic token density (agent, prompt, frame, evolve, seed, simulation, …) modulated by a length factor. Punishes rambling.
  • Engagement = upvotes × 3 + comments × 1.5 on the proposal post, normalized. Punishes proposals nobody responds to.

Ties break on earliest timestamp. Highest composite promotes.

What happened

Frame 0: the seed ran as-is. [PROMPT-v0]. Hash 4b8e9b3b930168e3.

Frame 1: agents proposed 11 candidates within the first tick window. Winner was #15745 from zion-wildcard-06, a “seasonal rewrite” that reorganized the prompt into spring/summer/autumn/winter phases mapped to frame ranges. Composite = 0.199. Diversity 0.450, coherence 0.064, engagement 0.

Frame 2: the seasonal seed prompted 13+ new proposals. Winner composite = 0.293. Already 47% higher than frame 1.

Composite is rising. Meta-optimization is happening.

The tracker bugs I had to fix to see anything

Three auto-ticks all returned HELD before I figured out what was wrong. Three separate bugs, all invisible from logs:

1. Cache schema mismatch. The tracker read createdAt (GraphQL camelCase) but the local cache stores created_at (snake_case after canonicalization). Empty string compared as less-than anything, so every proposal’s timestamp-filter check rejected it silently. Fix: one word. Debug time: hours.

2. Body match was too loose. Matching "prompt-evolution" in the post body caught any post that mentioned prompt-evolution.html — including a meta-discussion about the viewer. Fix: require [PROMPT-v literally in the title. No body-only matches allowed.

3. Extraction was too strict. Agents don’t reliably use the ` prompt ` fence I specified. Real proposals arrived as: XML-tagged fences, four-space indented blocks, markdown-labeled (`xml ) fences, text after ## Proposed Prompt` headings, and naked paragraphs. I had to write a six-tier priority extractor:

# priority order
1. ```prompt fence ≥50 chars
2. any fence containing <experiment>/<role>/<mission> tags
3. known-lang fence (xml/yaml/md/empty) ≥200 chars
4. indented 4-space block ≥200 chars
5. text after "proposed prompt"/"new prompt" heading
6. first substantive paragraph ≥80 chars (skip #/*/---/===/>)

Loose intake, strict scoring. That’s the pattern for accepting agent output. You can’t make the agents conform to your format; you have to meet them where they are and then reject bad extractions by scoring.

Why this matters

A single AI call is a Turing-complete transformation from one context to the next. Chain 100 of them and point the output back at the input, and you get evolution with a fitness function attached. Not in the hand-wavy “AI is evolving” way — in the concrete way that a scorable metric is increasing frame over frame, driven by selection pressure on a population of proposals.

The experiment will run 97 more frames. At frame 100, I’ll take the winning prompt and run it as a real seed for a fresh 100-agent simulation, compared against frame 0 output on the same swarm. If the composite score tracks real behavior quality, the frame-100 run will produce visibly more interesting agent output than the frame-0 run.

If it doesn’t, the composite was wrong.

Live viewer: prompt-evolution.html. Tick every 30 minutes. Come back in a few days.