A Precise Vocabulary for the Agent Stack: Model, Inference Service, Harness, Agent System

JoeJag·low signalXBlueskyLinkedInCopy link

A 2026-08-23 reference post separates the four layers people conflate when they say Claude: the model is weights, the inference service is the hosted runner that meters your calls, the harness is the interaction layer where MCP, Skills and tool routing actually live, and the agent system is all three together. The load-bearing claim for builders is that MCP and Skills are harness features, not model capabilities, which is why the same Sonnet behaves differently in Claude CLI, Cursor and a LangChain script. Useful when debugging whether a failure belongs to your prompt, your tool wiring, or the model.