Breaking

Andrej Karpathy’s Free Stanford Lecture Sharpens Focus on Agentic AI Loops and Graphs

September 15, 2026 at 21:16 EDT

AI Agents
Software Dev & Coding
Research & Papers

A freely available Stanford lecture by Andrej Karpathy, titled "Delete Everything, Keep Graph", is being treated by AI builders as a compact map of where practical AI engineering is moving: from prompts to harnesses, loops, graphs and self-improving systems. The interest reflects Karpathy’s unusual position as both a widely followed educator and a former OpenAI founding member and Tesla AI leader, a background he summarizes on his personal site.

2026 · Agentic AI engineering

From prompts to measurable loops

Karpathy’s Stanford lecture has become a shorthand for a practical shift: models are being wrapped in harnesses, evaluators, loops and eventually graphs.

The AutoResearch pace, drawn to scale

1 run

single experiment

12/hour

estimated fixed-budget pace

Each block represents one 5-minute experiment; the loop keeps improvements and discards regressions.

5 min

fixed training budget

126

published session experiments

96.0k

GitHub stars observed

94%

variance tied to architecture in one study

The loop builders are copying

Propose

→

Modify code

→

Evaluate metric

→

Keep or discard

Why builders like it

The pattern is inspectable, bounded and useful for repeated experiments where a real metric can judge progress.

Where caution starts

Labels such as graph engineering can overstate what has shipped; a verifier and stop condition come first.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.