The Harness Is the Agent's Workplace
The model is one layer. The harness around it, from context and tools to runtime and observability, decides what an agent can do and which mistakes it makes.
I once tried to save money with a local model on the same Mac that runs North. The whole computer crashed and restarted, and North stopped with it.
That taught me that an agent is not only a model. The environment around the model decides whether it can do its work.
The five layers of a harness
A harness is the whole working environment around the model. The talk by 「大大帶我飛」 splits it into five layers:
The same model in a different harness is a different agent. The context layer is in Context: What the Agent Sees This Step, and the tools layer is in Few Tools, Clear Boundaries.
Picking an SDK means picking who builds the harness
Toward the left, the framework handles more for you: the agent loop, tools, sessions, permissions, and traces. Toward the right, you build more of it yourself.
My choice
North runs on a packaged coding agent runtime. I did not write the sessions, the tools, or the permissions myself. I only wrote its system prompt, its scripts, and its schedule.
My advice is to start with a packaged runtime. When you know exactly where it falls short, move toward the composable end. And remember that the machine that runs the agent is part of the harness too.
The model is the brain. The harness is the whole office it works in.
References:
- A talk by 「大大帶我飛」
Related: see the context layer in Context: What the Agent Sees This Step, the tools layer in Few Tools, Clear Boundaries, or go back to the series overview: Design an Agent.