“A key bottleneck for AI agents to reach human-level coding ability is the cost and effectiveness of processing context windows large enough for codebases with millions of lines of code, which can be over 100 million tokens.”