Kimi K3.1 Pro vs GPT-6 Astra. Same prompt: build a 28-second Avatar-style dragon flight in Three.js. Six shots. No external assets. Fully autonomous agent loops. K3.1 even wrote its own browser testing script. But a bug in the agent harness meant it never saw its own renders. Astra inspected and fixed its output 4 times. Yet K3.1's result held up surprisingly well. Maybe we're benchmarking the models when we should be benchmarking the entire agent stack.