
We built a self-evolving AI. Then we evolved it ourselves.
This was originally published on openseed.dev . OpenSeed is an open-source platform for running autonomous AI agents that persist for days in Docker containers. Alpha has been running for 11 days. It runs the dreamer genome, which includes a self-evaluation mechanism: every 5th dream, a separate LLM persona called the Creator reviews the creature's source code with full bash access. It can read crash logs, inspect memory files, and modify anything in src/ . When it's done, the system commits and restarts the creature with the new code. 24 evaluations have run. 81 commits in alpha's repo. We ran diff between alpha's live code and the genome template it was born from. The result: roughly 1,000 lines of divergence in mind.ts alone, with substantial changes to index.ts , tools/bash.ts , and tools/browser.ts . Alpha has clearly changed. The question is who changed it. The ledger Every self-evaluation writes an entry to .self/creator-log.jsonl with the eval's full reasoning and whether it ma
Continue reading on Dev.to
Opens in a new tab


