SwarmLab

A monorepo of weird agent-swarm experiments — each one a wind-tunnel test of a faculty in the heybeaux stack (Parliament, Lattice, Sonder, Engram, ACR/AWM) — plus a Svelte observatory to navigate, run, and learn from them.

The point is learning, not shipping product. Weird is a feature. Emergent behavior you didn't predict is the win condition.

The Prime Rule

Only agents write the code. Every line in core/, observatory/, and every experiment is written, reviewed, and committed by an autonomous agent team. A human writes the charter and the specs; agents build. This isn't a gimmick — it forces the operator to learn orchestration, spec-writing, and reading emergent behavior instead of implementation. That's the real lesson.

The constitution the team builds under is TEAM.md. The running log of what each builder learned is JOURNAL.md — read it; the findings are the product.

Layout

swarmlab/
  core/          shared substrate — spawn, bus, score, trace (library only)
  observatory/   Svelte 5 dashboard — navigate experiments, replay traces visually
  experiments/   one folder per experiment, each built on core/
  specs/         detailed specs written before a build
  docs/          handovers, screenshots, and the North Star roadmap
  TEAM.md        the team charter / constitution
  JOURNAL.md     chronological learning notes (one paragraph per unit)
  SYNTHESIS.md   cross-experiment findings and stack recommendations
  BRAINSTORM.md  the canonical 12-idea brief

The substrate: core/

Every experiment depends on four contracts (core/src/types.ts):

• TraceEvent — spawn | message | score | kill. Everything an experiment does is one of these, timestamped.
• RunRecord — an experiment run: an ordered list of events, persisted as JSONL.
• AgentSpec — id + system prompt + optional model/context.
• AgentHandle — send / onMessage / kill.

Modules: spawn (agents, with a pluggable AgentRuntime seam — StubRuntime for deterministic in-process runs, real LLM backends plug in behind the same interface), bus (message passing), score (fitness/observation), trace (record + deterministic replay()). See core/README.md for the full API.

The showpiece: observatory/

A hand-scaffolded SvelteKit + Svelte 5 (runes) app. Three views: a lab index grid reading real experiments/* folders, an experiment detail page rendering each README with per-run stats, and the hero trace replay — agents as glowing nodes, messages flying the bus with topic labels, score events pulsing rings, kills graying nodes, all on a virtual timeline with play/pause/speed/scrubber.

cd observatory && npm install && npm run dev

Experiments

Each experiment tests a faculty. NN is the brainstorm idea number, not build order.

#	Experiment	Tests	What it probes
01	Telephone Compiler	Sonder / ACR	spec→code→spec drift over 10 hops
02	Adversarial Pair	Lattice / AWM	coder vs test-writer alternating on one file
03	Prompt Darwinism	ACR / AWM	mutate prompts, score, breed winners
04	Consensus Under Lies	Parliament / Lattice	N agree; 1-2 secretly mislead
05	Bug Telephone	Lattice	injected bug survives a reviewer chain
06	Self-Modifying Swarm	Lattice	agents rewrite each other's prompts
07	Minimal Language	Sonder	swarm designs a DSL + interpreter blind
08	Rumor Mill	Engram	gossip/epidemic propagation on a mesh
09	Overnight Cathedral	Lattice / Sonder	long-horizon iterative build, each commit reviewed
10	Economic Agents	Lattice / AWM	token-budget scarcity shapes collaboration
11	Reverse Engineer	Sonder	black-box behavioral reconstruction
12	Schema Negotiation	Sonder	two agents negotiate a shared protocol
13	Team's Choice	—	the team invents its own experiment

Each experiment folder ships: source built on core/, committed runs/*.jsonl traces, a README.md explaining what it tests and what was actually observed, and a note in JOURNAL.md. Every result is real — a red trace beats a faked green one.

Evidence and roadmap

• SYNTHESIS.md is the cross-experiment stack intelligence report.
• CLAIMS.json is the machine-readable claims ledger: each RT headline maps to run ids, trace paths, score fields, reproduction commands, and stack recommendations.
• docs/EVIDENCE-LEDGER.md explains the ledger format, verification command, and run metadata convention for new traces.
• docs/NORTH-STAR-ROADMAP.md formalizes the next phase: evidence supply-chain hardening plus specs 22–30.
• Raw traces are intentionally committed under experiments/*/runs/*.jsonl; they are the replayable evidence corpus behind the writeups.

Evidence and replay

npm run verify:evidence

The verifier loads CLAIMS.json, replays every referenced JSONL trace with the shared trace reader, and asserts the listed score fields. This verifies the committed evidence corpus; use the per-claim reproductionCommand when you need to regenerate fresh traces.

Running an experiment

npm install
npm run typecheck        # strict; no `any` in public signatures
# each experiment exposes its own run entry — see its README

Traces land in experiments/NN-*/runs/ and are read back deterministically by replay(), then visualized in the observatory.

Status

Built by the autonomous team, in order: core/ → observatory/ → experiments 01–34, including the Aegis retest chain through RT-25. The machine ledger currently contains 25 verified claims backed by 51 admitted traces and 490 exact score assertions. Progress is committed after every green unit; check JOURNAL.md, SYNTHESIS.md, and the claims ledger for the findings and current capability boundaries.