Developmental Software Ecology

An experimental platform for studying executable cultural evolution in persistent agent ecologies.

The project asks whether initially identical, self-directed LLM agents with bounded private memory can accumulate measurable capability across agent turnover through persistent, executable Git artifacts.

Current status

V0.1 — M0: Deterministic Empty World

The first milestone intentionally contains no LLM, Forgejo, sandbox, embeddings, or dashboard. M0 establishes the deterministic experimental core:

• immutable experiment manifests
• identical initial agents
• monotonic event ordering
• reducer-based state transitions
• snapshots
• replay
• canonical state hashing

M0 gate

manifest
  -> world creation
  -> 5 identical agents
  -> deterministic event stream
  -> snapshot
  -> shutdown
  -> restore
  -> replay
  -> identical state hash

Research framing

This is not a generic "AI town" or human-life simulator. V0.1 tests a narrower falsifiable question: whether persistent executable artifacts can function as cumulative culture across individual agent turnover without inheriting private episodic memory.

Development workflow

Implementation work proceeds on feature branches and is merged through pull requests after deterministic replay tests pass.