Descriptions:
Nick Saraev demonstrates how to apply Andrej Karpathy’s open-source “auto-research” framework — originally built to let AI models autonomously run machine learning experiments overnight — to real business optimization problems using Claude Code. The pattern is simple: an AI agent modifies something, measures an objective metric, logs what it learned, and repeats in a continuous loop.
Saraev’s primary example is a cold email optimizer. He built a folder called `email_optimizer` containing an `orchestrator.py` file that runs every four hours via GitHub Actions. The agent generates challenger email copy variants, tests them against a baseline using reply rate as the objective metric, and writes learnings to a `resource.md` file that informs the next generation of copy. Over successive cycles, the system builds a compounding knowledge base of what works — self-improving without human involvement.
The video walks through the complete setup: cloning Karpathy’s repo from GitHub, writing a `program.md` file that specifies a goal, a measurable metric, and a test method, and deploying to GitHub Actions or Modal for scheduled execution. Saraev emphasizes that the pattern generalizes to any workflow with a quantifiable output — landing page conversion rates, ad copy CTR, onboarding completion — making it a broadly applicable architecture for autonomous business optimization using Claude Code as the orchestration layer.
📺 Source: Nick Saraev · Published March 12, 2026
🏷️ Format: Hands On Build
