In the swiftly changing world of AI, confusion often arises from the rapid evolution of its vocabulary. Terms like 'harness' and 'scaffold' have gained prominence, yet their meanings differ widely across various contexts. This inconsistency can overwhelm both newcomers and seasoned practitioners as they struggle to keep up with the latest developments.
Following the International Conference on Learning Representations (ICLR) in 2026, a pertinent question emerged from attendees: "What do you mean by the terms 'harness' and 'scaffold' in the context of agents? I have heard a lot of explanations while I was at ICLR, but I could not understand why they did not converge to a single explanation." This query captures the current state of discourse among AI professionals, underscoring the urgent need for clarity.
To tackle this confusion, a new glossary has been introduced, focusing on the most frequently misused and misunderstood terms in the AI agents domain. The goal is not to create an exhaustive dictionary, but to ground discussions around concepts often mixed up or assumed to be self-explanatory. This glossary is particularly relevant for those involved in building or deploying AI agents, as well as users of platforms like Claude Code, Codex, or Hermes Agent.
The Role of Models, Harnesses, and Scaffolds
At the core of AI agents is the large language model (LLM), which processes input text and generates output. However, models such as Claude, Qwen, or GPT operate independently, lacking memory between calls or continuous loops. To transform a mere LLM into a functioning agent, it requires a 'harness'—a framework that enables execution by wrapping the model and providing essential context management.
The 'harness' consists of various elements that are not part of the LLM itself, including system prompts, tool descriptions, and mechanisms that manage contextual information across steps. This behavior-defining layer significantly influences how the model interacts with the world, whether during training or inference. As highlighted in Claude Code's documentation, "Claude Code serves as the agentic harness around Claude."
Conversely, the term 'scaffold' often describes the underlying infrastructure that supports the harness. This includes hooks, runtime configurations, and even directory structures that the harness relies upon. Understanding the distinction between these two terms is key, particularly when navigating training pipelines where these components are reasoned about separately.
Diverse Applications and Implementations
Different AI products exhibit varying levels of coupling between their harnesses and the models they employ. For example, Claude Code and Codex are closely integrated with their respective providers' models, while alternatives like Antigravity CLI and Hermes Agent offer flexibility by allowing users to plug in any model of their choice. This diversity reflects the evolving nature of AI, where terminology and application can shift unexpectedly.
The glossary aims to bridge the gap in understanding these terms, providing a practical mental model that enhances discussions within the AI community. As the field continues to grow, establishing a shared vocabulary will be important for clearer communication and advancing collaborative efforts.
In a field characterized by rapid advancements and shifting narratives, clarity is essential. As AI agents become increasingly integrated into various applications and industries, a consistent understanding of foundational terminology will be key for both developers and users. The glossary represents a proactive step toward achieving this clarity, paving the way for more informed discussions and innovations in the AI space.
Quick answers
What is the difference between harness and scaffold in AI agents?
The harness encompasses the elements that facilitate the model's execution, while the scaffold refers to the infrastructure that supports the harness.
Why is there confusion around AI agent terminology?
Rapid evolution in the field has led to terms being reused and redefined, resulting in a lack of consistent understanding among practitioners.
What is the purpose of the new glossary?
The glossary aims to clarify frequently misused terms in AI agents to facilitate better communication and understanding within the community.
