emergent

emergent

An Agent Skill for Claude Code that runs an autonomous ML research loop to hunt for emergent properties — capabilities that arise from training, scale, or composition but were never explicitly designed. Heavy compute runs on a remote RunPod GPU over SSH; your context window stays small while the GPU does the work.

The goal is to surface genuinely novel phenomena with the skepticism of a real scientist: every surprise is a candidate, not a discovery, until it survives reproduction, ablation, and a hunt for the boring explanation.

What it does

1. Explores your local project and the RunPod environment in parallel.
2. Writes an emergence.json config (probes, run command, budget cap).
3. Runs a baseline, then loops: orient → hypothesize → run → measure → verify.
4. Grounds hypotheses in arXiv literature via the arxiv CLI.
5. Keeps reproduced findings, discards noise, logs everything to a research journal.
6. Respects a spend/time budget and stops the pod when done.

Install

npx skills add tylergibbs1/emergent

Or install manually:

cp -r skills/discovering-emergence ~/.claude/skills/discovering-emergence

Or clone directly into your skills directory:

git clone https://github.com/tylergibbs1/emergent.git \
  ~/.claude/skills/emergent

Then in Claude Code:

find emergent properties — here's my runpod: ssh root@<ip> -p <port> -i ~/.ssh/id_ed25519

The skill triggers on phrases like "find emergent properties", "emergence research", "autonomous ML experiments", or when you provide a RunPod for experimentation.

Layout

skills/discovering-emergence/
├── SKILL.md                  # main loop and workflow
├── references/
│   ├── EMERGENCE.md          # what counts as emergence + verification checklist
│   ├── RUNPOD.md             # SSH ops, remote runs, budget tracking, pod shutdown
│   ├── STRATEGIES.md         # where to look when out of ideas
│   ├── LITERATURE.md         # grounding hypotheses with the arxiv CLI
│   ├── CONTEXT.md            # context-window discipline for long runs
│   └── CONFIG.md             # emergence.json field reference
└── examples/
    └── emergence.json        # sample config

Credits

Loop pattern adapted from autoresearchskill (inspired by Karpathy's autoresearch). Literature search via arxivcli. Plugin format follows linearprincipals.