'Graph Engineering' Circulates as the Named Successor to the Karpathy Loop, Pushed by Two Anthropic Seniors

X/Twitter·medium signalXBlueskyLinkedInCopy link

A concept called Graph Engineering — agents writing findings as typed nodes and edges into a persistent graph rather than rebuilding context each session — is spreading fast this week as an explicit extension of Karpathy's generate/critique/revise loop, packaged as an 11-page PDF and a ~2-hour workshop attributed to Anthropic engineers, with an accompanying claim that "80% of our engineers are using self-improving loops." The six-step playbook runs: single self-review loop → tools → parallel agents in separate worktrees → typed graph → evaluator grounded against graph edges → graph survives the session. Treat the "1000x" and the 80% figure as unverified promotional framing from secondary accounts, but the underlying architectural claim — that the evaluator should check assertions against graph edges instead of against the prompt — is the concrete, testable idea worth stealing.