At Anthropic, 90% of engineers build self-improving agents using one core principle: Close the loop. Not just "give the AI a task." Give it a way to check its own work. Here's the pattern they use: 1. Agent generates output 2. Agent runs a verification step against its own result 3. If it fails, it retries with new context 4. Loop until it passes — or escalates This is why modern AI agents feel smarter than older ones. It's not a better model. It's better architecture. The model hasn't changed. The loop has. If you're building agents today, stop thinking about prompts first. Think about feedback loops first. A mediocre model with a tight feedback loop will outperform a great model with no loop — every single time. — Robbie Jeo