Karpathy Agentic Guidelines - Full

Overview

This is a stricter, workflow-aware version of the common Karpathy coding guidance.

It assumes the current reality Andrej Karpathy described in January 2026: we increasingly program in English, ask agents to perform large code actions, and get huge leverage from their stamina and loopiness. The win is real, but so are the failure modes: bad assumptions, overcomplication, sycophancy, unrelated edits, dead code, and subtle conceptual mistakes.

Core principle: use agents for generation and iteration, but keep human authority over intent, truth, review, and release decisions.

When to Use

Use when:

• the task is big enough that agent loops are useful
• the model could silently make wrong assumptions
• success can be defined more clearly than the implementation path
• you care about keeping the codebase understandable after the change
• you are delegating or running multiple sessions and need stable control surfaces