Harness Routing Principles: Thick vs Thin Harness, Early Rejection, and Avoiding Unnecessary LLM Calls

The article, the third part of a tutorial on harnesses, focuses on routing in AI agent systems. It contrasts a 'thick' harness, which helps weaker models by adding entities like Intent, Route, Shortcut, Capability, and ContextRequirement, with a 'thin' harness, which relies on strong models and emphasizes Capability, Workflow, ToolContract, Policy, Approval, Task, and Trace. It argues that the thin harness should be used when possible, but the thick harness can compensate for model weaknesses without compromising security boundaries. A key principle is to reject unexecutable requests as early as possible, through steps like input normalization, context binding, access checks, workflow selection, and policy gates. The article also stresses not using LLMs when not needed, handling structured events or exact commands with code instead of models, to save resources and reduce errors. It mentions that the choice of harness thickness is the fundamental design question, and that thick harnesses can be gradually 'thinned' by moving intent recognition to the model.