FlowtandBriefAI · Software · ResearchMeet Flowtand

Hugging Face Glossary Separates Harness, Scaffold in AI Agent Terms

Announcement

Hugging Face

Browse Agents topic page →

A Hugging Face blog post argues that AI agent vocabulary has drifted, so it proposes a practical mental model distinguishing the model, scaffolding, harness, context engineering, tools, skills, sub-agents and training concepts. Its central move is to treat the harness as the execution loop and the scaffold as the model's instructions, tools and formatting — a split the authors say matters most when reasoning about training pipelines.

What the glossary proposes

Published on the Hugging Face blog and dated May 25, 2026, the post by Sergio Paniego and Aritra Roy Gosthipaty responds to confusion the authors observed around terms like 'harness' and 'scaffold' after ICLR 2026. Rather than introducing a tool, it offers a mental model. The model is the LLM itself: text in, text out, no memory between calls and no loop. Scaffolding is the behavior-defining layer — system prompt, tool descriptions, response parsing and context management. The harness is the execution layer: it calls the model, routes tool calls and decides when to stop. The authors note that products such as Claude Code, Codex and Antigravity CLI call the whole surrounding system a harness, and that Claude Code's own documentation describes it as an agentic harness around Claude. The distinction, they argue, matters most when the two need reasoning about separately, as in training. An agent is framed as model plus harness, and an orchestrator as a higher-level controller managing agents that each run their own harness.

A documentation scenario

A technical writer or developer advocate maintaining onboarding material for a coding-agent product could use this split to structure a glossary page. A conditional approach: define the model, then the scaffold (prompts, tool schemas, output format), then the harness (the loop, tool dispatch, stop conditions), and finally the product itself. The post stresses that the model, the harness and the product are three different things, which gives the writer a concrete way to explain why two products on the same model can feel different. This applies to explaining vocabulary, not to asserting how any specific product behaves.

Trying the vocabulary on a real project

A team could trial the scaffold/harness distinction on its own agent codebase. Start by labelling each component: which files hold instructions, tool descriptions and context handling (scaffold), and which hold the loop, tool routing and stop logic (harness). Baseline the current setup by recording, for a fixed set of representative tasks, where failures occur and whether they trace to instructions or to loop control. Then make one change in one layer only — for example adjusting a tool description or a stop condition — and rerun the same tasks. Success means you can attribute a difference to the layer you changed. Failure checks include ambiguous component boundaries where a change spans both layers, or tasks too few to distinguish noise from effect. Keep a human reviewer deciding whether an observed difference is meaningful.

Trade-offs and when not to use it

The authors are explicit that these definitions are not universal and that frameworks reuse words differently, so this vocabulary is a discussion aid rather than a standard to enforce. Adopting the narrow scaffold/harness split may conflict with vendor documentation that uses 'harness' broadly; teams should expect to translate rather than replace product language. The post also notes that the line between tool, skill and sub-agent shifts across frameworks, so precise boundaries are not guaranteed. A team should avoid imposing this split when communicating with stakeholders already fluent in a vendor's own terminology, or when the distinction adds documentation overhead without a decision it informs.

Context and limitations

  • Based on a short source excerpt; the full article was not analyzed.
  • The authors state many terms lack universally accepted definitions and frameworks use the same word differently.
  • The post is a glossary proposing a mental model, not a report of tested system behavior or benchmark results.
  • Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.