Golfing ARC-AGI with a Evolutionary Agentic Loop

How far can you push LLM-driven program synthesis in a real-world code golf competition? In this deep dive, I share lessons from building a lean, agentic system that tackles the NeurIPS 2025 Google Code Golf Championship and solves hundreds of ARC-AGI tasks. You’ll see the architecture in action—from prompt engineering and evolutionary retrieval to hard-won code-golf templates and context-rot hacks. Along the way, I highlight what worked, what broke, and how a handful of manual breakthroughs translated into reusable patterns for automated byte minimization. Whether you’re interested in LLMs, agent design, or just love a good code golf story, you’ll find practical takeaways and hard-earned war stories here.