Raw LLM Responses

It's new, still working out the kinks, but tbh, stuck in a loop, I've yet to see. We have a well-structured planning system. Tokens, I use Claude Code Max 20, so I don't really have to worry about tokens, but we do consider them. We're always looking at ways to reduce tokens. Watchdog is a good example, rather than have a detailed prompt instruction, it's now a program, so it reduces token usage on that end. We try to give agents just enough awareness to know that something exists — tools, a process, etc. When an agent interacts with email, it has general command knowledge, but how to structure your send or response context is only presented when the tool is being used. Drone is our command routing. Agents don't need to know 1000 commands, paths — drone @agent <args> resolves it all for the agents, so that alone is a massive token saver. I also have an AI language compression system in the works, AIPL, AIPass Language. For AI-to-AI and AI files, rarely or never looked at by a human. Seeing around 7-40% token reduction in daily activities. It's still human readable, but more suited for AI, so if you read your agent memories, you'll understand but be like 'not for me.' Haha — and that's the whole point. Your outputs will be totally fine