firehose> #llmops

Harness / Model Separation

The agent harness and the model that drives it are separable artifacts. The harness is everything that is not weights: the system prompt, the tool definitions, the agent loop, the permission model, the file conventions, the subagent machinery. The model is an interchangeable inference backend behind them. The source's analogy — Claude Code is the car; the chat model is the engine — is doing real work, because the two are usually bought together and therefore usually reasoned about together. Pulling them apart reframes what you are paying for: when an operator hits a "context limit" or a "token limit," that is a property of the model contract they bought, not of the harness they are using. The harness itself is a piece of software running on their machine, and it will point at whatever inference endpoint they tell it to (see BYO Inference Endpoint).

This separation is what makes Model-Tier Routing mechanically possible at all — you cannot route a sub-role to a cheaper tier unless the loop around that sub-role is indifferent to which model answers. It is also why the harness, not the model, is increasingly where vendors differentiate: the loop is reproducible, but the accumulated tool design (Agent-Computer Interface (ACI)), prompt scaffolding, and conventions are the durable artifact.

Separability is not free, and the failure mode has its own page: a harness encodes an implicit contract that the substituted model must satisfy — see Harness / Model Fit. The engine bolts in; the car may still drive badly.

Claims


Linked from