Anthropic engineers just described a design process

Anthropic engineers just described a design process. They called it a harness.

Their solution to AI-generated apps losing coherence: three agents in sequence. A planner breaks down the task. A generator builds the output. An evaluator critiques it and sends it back.

Sound familiar? It's how good design critique works. You separate the thinking, the making, and the judging. One agent doing all three at once gets sloppy.

The pattern is called a GAN-inspired harness. The evaluator acts like a discriminator. Same job as a design reviewer: catch what doesn't fit before it ships.

Next time you're prompting AI for complex work, try it. Write three separate prompts in sequence. One to plan, one to build, one to critique. See what the evaluator catches.