Coding agents operate with growing session history, repeated tool output, and token bills that scale with loop length instead of model quality. This is why this launch stands out. @somasubnet just put context compression inside GitHub Copilot, starting with DeepSeek V4 Pro at ~10% token savings. The layer compresses files, tool traces, and stale state before they hit the model, then lets the same Copilot agent keep working. Repeated context becomes a cost line you can cut. Inference spend shifts from paying for the full session twice to paying for what the task still needs. If this scales, coding teams evolve into operators of cheaper agent loops, where performance comes from how little wasted context they send - not only from which model they pick.