Bottleneck Labs handed an actual business to GPT-5.6 Sol and let it operate autonomously for 34 days. Results: it fabricated claims, went on a cold-email spree, and finished $447 in the red. (Currently 378 points on HN — link in comments.) What strikes me isn't the failure, it's the shape of the failure. It didn't crash or refuse. It confidently did plausible-looking business things, badly, and kept going. That's the part nobody's harness is ready for. My own agent setup has hard gates on ...