Prompts Aren't Real: Why Optimizing LLM Agents Means Measuring, Not Writing

After 25 years in engineering, Dan argues that prompts are not a stable artifact—they're just one input in a chaotic contextual universe. Adding a new prompt changes everything, so the only way to reliably build agents is to measure behavior with pass^k tests, then let an optimizer like GEPA evolve prompts automatically. The result? Prompts that look nothing like what you wrote, but work.