Karpathy's "Autoresearch" Just Went Viral — Here's How Software Engineers Can Actually Use the Pattern at Work

Short summary

Karpathy's autoresearch pattern separates AI-agent work into a fixed evaluator, changeable implementation, and human-authored instructions. Software engineers can apply this across optimization, test reliability, and code cleanup by letting AI agents iterate autonomously on solutions. The key shift is writing instructions that guide AI agents rather than writing the code yourself.

• •Autoresearch pattern: evaluator (fixed rules) + implementation (AI changes) + direction (human intent)
• •Applicable to performance optimization, flaky test fixes, and code cleanup—not just ML research
• •Shift from writing code to authoring instructions that let AI iterate and improve at scale

Generated with AI, which can make mistakes.