Graph Engineering: The Karpathy Loop, Improved 1000x by Itself — The Anthropic Playbook
This independently compiled July 2026 technical note connects Andrej Karpathy’s autoresearch and AgentHub sketches with Anthropic workflow and knowledge-graph patterns to propose an incremental architecture for verifiable, collaborative, provenance-aware agent systems. The source title’s ‘1000x’ is rhetorical, not a measured improvement claim; the note separately discusses workflows that can schedule up to 1,000 agents.
4 sections · 12 lessons
Course outline
Measured Loops
- Externalize the Bottleneck
- Program the Program
- Build a Safe Ratchet
- Bound an Extraction Loop
Collaborative Search
- Preserve Experiment Lineage
- Coordinate Without Transcripts
- Generate and Reduce Work
- Design a Migration Workflow
Durable Graphs
- Extract Typed Claims
- Resolve Without Corruption
- Query Bounded Evidence Paths
- Investigate an Incident Path
Production Discipline
- Join Lineage and Provenance
- Separate the Five Planes
- Budget, Stop, and Evaluate
- Approve a Graph-Grounded Swarm