“Sholto Douglas predicts that by the end of 2025, there will be conclusive evidence of long-running agentic performance from software engineering agents capable of doing real work.”