Anthropic reveals GAN-inspired coding agent harness

The Harness architecture signals a shift from "chat-to-code" to autonomous engineering systems where orchestration logic is as critical as the model. By separating creation from evaluation, Anthropic solves the "self-evaluation bias" that plagues single-agent systems. The system uses a GAN-inspired feedback loop where a Generator is pitted against a skeptical Evaluator using Playwright for live UI/API verification. Specialized agents operate in fresh context windows, using Git and progress logs for state handoff to mitigate context decay. Benchmarks show the harness delivers polished, functional apps in 6-hour runs ($200 cost) that solo models fail to produce in 20 minutes. Evolution in models like Opus 4.6 is simplifying the harness by removing sprint-level decomposition while maintaining the essential evaluation layer.