Harness vs. Scaffold: Getting AI Agent Vocabulary Right

The AI agent field is awash in overlapping terminology—’harness’ and ‘scaffold’ are often used interchangeably, causing confusion even among practitioners at top conferences. This glossary tackles the core tension: terms evolve faster than shared understanding, making it hard for newcomers and experienced builders alike to reason about agent architectures precisely.

The article draws a clean distinction: scaffolding is the behavior-defining layer (system prompt, tool descriptions, context management) that shapes how the model sees the world, while the harness is the execution layer that calls the model, handles tool calls, and decides when to stop. Products like Claude Code and Codex often call the whole non-model stack a harness, but separating the two is crucial for training pipelines and system design. Other clarified terms include context engineering, policy, tool use, skills, and sub-agents, each grounded in how they function in real agent loops.

For builders, the takeaway is practical: don’t conflate the model, the harness, and the product—they are three different things with different optimization criteria. Using precise vocabulary for scaffold vs. harness makes debugging, training, and cross-framework discussions far more efficient. The glossary is not prescriptive but provides a mental model that helps engineers reason about agent components independently.