AI Agents Need Shared Vocabulary to Scale
A glossary from Hugging Face clarifies terminology in the rapidly evolving AI agents field, distinguishing between key concepts like scaffolding, harness, and model that are often conflated or used inconsistently across frameworks and products. The authors argue that as the field matures faster than shared vocabulary, practitioners need precise definitions to communicate effectively about agent architecture and behavior. The piece covers foundational terms relevant to building, deploying, and training agents, acknowledging that many lack universally accepted definitions across different implementations.
TL;DR
- Scaffolding is the behavior-defining layer around a model: system prompts, tool descriptions, response parsing, and context management that shape how the model acts
- Harness is the execution layer that calls the model, handles tool calls, and decides when to stop, distinct from scaffolding though often conflated in product marketing
- Model refers to the LLM itself with no memory or loop between calls; it becomes an agent only when wrapped in scaffolding and a harness
- Terminology inconsistency across frameworks and products creates confusion for newcomers and practitioners, motivating this attempt to establish practical mental models
Why It Matters
AI agents are moving from research into production, but the field lacks standardized vocabulary for core architectural components. Miscommunication about scaffolding versus harness, or what constitutes an agent, can lead to misaligned expectations between teams building, deploying, or evaluating these systems. Clear definitions become critical as the field scales and more practitioners need to collaborate across different frameworks and implementations.
Business Impact
Organizations deploying AI agents need shared language to specify requirements, evaluate vendor claims, and build effective teams. Products like Claude Code and Codex use 'harness' to describe their entire agentic layer, while others distinguish between scaffolding and harness, creating confusion in procurement and integration decisions. Standardizing these terms reduces implementation friction and helps teams accurately assess whether a tool meets their needs.
Key Implications
- Vendor marketing often obscures architectural distinctions by using 'harness' as a catch-all term, making it harder to evaluate what is actually customizable or portable across models
- Training pipelines require separating scaffolding from harness to reason about them independently, suggesting that production systems will increasingly need this distinction explicit in their architecture
- The lack of consensus definitions means teams must establish internal glossaries to avoid miscommunication, adding overhead to agent development and deployment projects
What to Watch
Monitor whether the AI agents community converges on these definitions or develops competing standards. Watch for how major frameworks like LangChain, AutoGen, and vendor-specific tools (Claude Code, Codex) evolve their documentation and whether they adopt or resist these distinctions. The pace of convergence will indicate whether the field is maturing toward standardization or fragmenting further.
Subscribe to the newsletter
The latest stories and analysis, delivered to your inbox.
Free. No spam. Unsubscribe any time.
