11-page 'Graph Engineering' PDF maps Karpathy's agent loop to Anthropic's knowledge-graph playbook
mdancho84 · x · 2026-10-21
An 11-page PDF titled Graph Engineering: The Karpathy Loop, Improved 1000x by Itself is circulating on X. Caveat: it is independently compiled (July 2026) from Anthropic's Knowledge Graph Construction Cookbook and Karpathy's public courses—not official or endorsed.
Key points:
- Karpathy's autoresearch is a single-agent loop (700 ML experiments in 2 days, 20 optimizations); AgentHub scales agent swarms via a shared commit graph/DAG to 1,000 parallel sub-agents without orchestrator context growth.
- Anthropic's Dynamic Workflows splice fresh sub-agents in parallel; the Cookbook replaces NLP pipelines with structured-prompt entity extraction.
- The staged build path: measured loop → tools and planning → multi-worker, building a typed graph only after specs stabilize.
- The knowledge graph acts as shared memory: assertions, audits, evals and versions become queryable graph objects.
More from coding & agent
- Dev builds phrok, an MCP server that lets grokbot make real phone calls — naxmax2019 · 2026-10-22
- Simulithic (YC F26) launches AI agents that run E2E tests like real users — KlausCodes · 2026-10-22
- echudhra.com: one-line curl/node/docker commands to test if your tools trust the leaked eMudhra keys — SuB8u · 2026-10-21
- Viral meme: every AI coding tool stands on the same Terminal Bench pillar — geoffwolfe · 2026-10-21
- Vercel's durable-harness moves coding agents to the cloud with auto-resume for month-long runs — mitsuhiko · 2026-10-21
- Cohere launches Rerank 5 to push AI retrieval precision, available now in Pro and Fast tiers — calabi_and_yau · 2026-10-21