Harness, Scaffold, and the AI Agent Terms Worth Getting Right

When a field evolves quickly, its vocabulary often evolves faster than its shared understanding. Terms start to blur, get reused in different contexts, or become shorthand for ideas that are never fully explained. We are currently seeing this happen in the field of AI Agents, where concepts are getting mixed together, some are renamed, and others are widely used for a few months before quietly disappearing.

This can be overwhelming for newcomers, and even for practitioners trying to keep up with the latest developments. After ICLR 2026, one of us ( @ariG23498 ) posted a question that captured this confusion well:

"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."

At evaluation time, the same pattern shows up as an eval harness : instead of collecting training data, it runs a fixed set of scenarios at a model checkpoint and records metrics rather than updating weights.

The term comes from reinforcement learning, where an agent is simply a function that takes an observation and returns an action. The environment takes that action and returns a new observation, and the loop repeats. That loop is still at the core of how LLM agents work.

A policy is the behavior an agent follows: given any situation, it defines the probability of taking each possible action. In LLM systems, part of that policy is learned in the model weights, but the behavior also depends on the surrounding scaffolding and harness. The same model can behave very differently depending on its prompts, tools, memory, and execution loop.

The terms above apply whether you're training or deploying. These four are specific to training, where the agent runs through tasks, gets scored, and its model's weights get updated. Every RL training system for LLMs is built around the same pipeline:

The execution layer inside the agent: it calls the model, handles its tool calls, decides when to stop. The harness is what makes the agent run. Scaffolding, defined above, is what the model works from: its instructions, its tools, its format.

Take a coding agent as a concrete example. The system prompt, tool descriptions, and the output format the model follows form the scaffolding. The loop that calls the model, handles its tool calls, and decides when to stop is the harness. At training time, the harness also runs many of these loops in parallel and feeds the results back to update the model.

How agents reach outside themselves: APIs, code interpreters, databases, web search, file systems. The model expresses the intent to use a tool in a structured format. Modern inference APIs surface this as a first-class object: the harness receives the call directly and routes it to the right function. The result gets fed back into context and the loop continues.

This glossary is our attempt to ground the terms that keep coming up without clear, consistent explanations. It is not meant to be a comprehensive dictionary of every term in the field. Instead, we focus on the concepts that are often mixed up, reused in different ways, or assumed to be obvious when they are not.

Most of these terms come up whether you're building an agent, deploying one, or just using tools like Claude Code, Codex, or Hermes Agent. The last section covers concepts specific to training models, which is more relevant if you work on that side of things.

Many of these terms don't have universally accepted definitions yet, and different frameworks use the same word differently. The goal here is not to enforce one correct vocabulary, but to provide a practical mental model that makes discussions easier to follow.