Karpathy's 700-experiment autoresearch loop plus Anthropic's knowledge graph cookbook: public yet ignored
blaizedsouza · x · 2026-11-22
A thread curates two public but overlooked resources that together reproduce Karpathy-style autonomous research loops:
- Anthropic's cookbook for turning documents into a knowledge graph with Claude, giving agents memory they can reason over
- Karpathy's autoresearch loop, which ran 700 experiments in 2 days and found 20 optimizations with no human steering
The reposter says wiring the graph half into his own setup immediately changed output quality: Claude stopped giving stock answers and actually worked through the problem.
The quoted long-form article, "300 Agents, One Graph, and a Loop That Edits the Loop," lays out the recommended build order for loops, graphs, dynamic workflows and routines — the author rebuilt the same agent setup four times this year and threw three versions away.
More from coding & agent
- Anthropic doc: one 'you own this codebase' line lifts Claude's code consolidation from 10% to 75% — lxfater · 2026-11-22
- Late-2026 agent checklist: A2A payments, 50-branch inference, self-building RAG — blaizedsouza · 2026-11-22
- Tip: convert Grok MCP routines to script-first jobs to stop wasting tokens — intellectronica · 2026-11-22
- Lovable x Paper MCP workflow: explore designs visually, then ship components — damienghader · 2026-11-22
- Redditor builds a 1:1 Earth-scale planet with Opus 5.5, one continuous shot from the Moon to a city — Promptmethus · 2026-11-22
- Agent engineering in four layers: open-source trueforge harness claims 2.7x token savings — blaizedsouza · 2026-11-22