A weighted answer to the recent wave of Copilot / coding-agent spend should be @somasubnet, especially with the new GitHub Copilot integration. Under the scope of SOMA, compression expresses what the agent still needs to finish the task, without sending the full session history every call. Thus, SOMA is pretty much like a smaller prompt, but at the agent-session level. SOMA has the power to: ▸ cut repeated files, tool outputs, and stale state ▸ keep the same model and the same Copilot workflow ▸ lower the token bill before inference And @somasubnet is first applying that layer to DeepSeek V4 Pro inside Copilot, with Early Access credits and more models next. Below is the live drop.