Agents that natively self-orchestrate, managing their own context, tools, and sub-agents, are the next big unlock in LLM performance. Right now, a skilled engineer building an optimized harness, with thoughtful data flow, separation of concerns, sub-agent management, etc., can make dramatic improvements over baseline for specific tasks. If a model could do this itself, that’d be a major step forward. You give it an objective and a set of tools, and it figures out the optimal way to orchestrate itself to do the task. For example, I’m building a very primitive AI scientist that I’ll open-source soon. Most of the work isn’t in the prompt, it’s in the harness… what the orchestrator sees, what sub‑agents see, what gets shared between them and when, where we summarize vs. pass raw data, and which tools each agent controls. Doing this allows me to dramatically improve what the model can do on its own. If a model can effectively design its own harness for a given problem, it’d be a huge step forward. My bet: self-orchestrating models… ones that manage their own context, tools, and sub-agents, will move the frontier almost as much as the jump from chatbot → reasoning did. Maybe more.

Just calling my shot here… I’m pretty confident in this. Someone can prototype this today (maybe I will!) by having a model write a harness for a given prompt in Python, slot that into a @daytonaio sandbox or something similar, and then passing the prompt to the harness.