---
title: "Harness glossary"
description: "The words that decide an agent setup: model, harness, agent, tool, skill, MCP, context window, sandbox, approval policy and headless mode."
canonical: "https://www.diginomads.win/harnesses/glossary"
source: "content/pages/harnesses/glossary.md"
updatedAt: "2026-11-10"
---

The agent market has a vocabulary problem: the same word means a product, a loop and a job title. This page fixes the meanings used across the site. It is a working glossary, not a standards document.

## The core four

| Term | What it means here |
|---|---|
| **Model** | The network that produces tokens. It has a context window, a price and a knowledge cut-off. It does not read your files on its own. |
| **Harness** | The program around the model: tools, permissions, session storage, context assembly and interface. Claude Code, Codex CLI and Aider are harnesses. |
| **Agent** | The model plus the harness plus the task loop. When someone says “the agent fixed the bug”, they mean the whole loop, not the weights. |
| **Tool** | A capability the harness exposes to the model: read a file, run a command, search the web, call an MCP server. A tool is a permission boundary as much as a function. |

A useful sentence: *the harness gives the model tools, and the agent is what happens when the model uses them on a task.*

## Context and memory

| Term | What it means here |
|---|---|
| **Context window** | The maximum amount of text the model can consider at once, including instructions, files, tool output and the conversation. |
| **Context assembly** | How the harness decides what enters the window: project instructions, file selection, search results, summaries. Two harnesses with the same model can assemble very different contexts. |
| **Compaction** | Summarising or dropping older context to stay inside the window. It is lossy; a decision that mattered can disappear. |
| **Session** | One conversation and its tool history. Resuming a session is not the same as restoring the repository state. |
| **Subagent** | A separate agent loop, often with its own context, that returns a result to the main session. Useful for research and review; not a substitute for a test. |

## Control and safety

| Term | What it means here |
|---|---|
| **Permission mode** | A preset that decides when the harness asks before acting: for example plan, ask, auto-edit or a fully permissive mode. |
| **Approval policy** | The rule for when the agent must stop and ask. It can be per tool, per path or per command pattern. |
| **Sandbox** | An operating-system or container boundary that limits what a command can touch. A permission prompt is not a sandbox. |
| **Allowlist / denylist** | Explicit rules for commands, paths, domains or tools. Prefer a short allowlist over a long denylist. |
| **YOLO mode** | A mode that skips approvals. It is a development convenience, not a security posture. |
| **Prompt injection** | Instructions hidden in content the agent reads, such as a web page or a file, that try to redirect the agent. Treat fetched content as data, not orders. |

## Extensibility

| Term | What it means here |
|---|---|
| **Skill** | A folder of instructions, references and optional scripts that an agent loads when a task matches its description. The open format is documented at [agentskills.io](https://agentskills.io/specification). |
| **MCP** | Model Context Protocol, an open protocol for connecting an agent to tools and data sources through a server. See [modelcontextprotocol.io](https://modelcontextprotocol.io/). |
| **Plugin** | A vendor-specific package that may bundle skills, commands, MCP servers and configuration. Portability depends on the vendor. |
| **Headless mode** | Running the harness without an interactive interface, usually for scripts and CI. The exact flag and output format differ by harness. |

## Workflow words

| Term | What it means here |
|---|---|
| **Bake-off** | Running the same small set of tasks through two or three harnesses on the same repository, then comparing with a scorecard. The [45-minute version](/harnesses/how-to-test-a-harness) is on this site. |
| **Diff** | The change between two repository states. A reviewable diff is small, has a reason and does not hide a deleted test. |
| **Reproduction** | The smallest command that shows the failure. A fix without a reproduction is a hypothesis. |
| **Rollback** | The documented way to return to the previous state. If it is not written down, it is not a rollback. |
| **Portable setup** | The instructions, prompts, skills and checks you own in plain files, independent of any one harness. See [a portable agent setup](/harnesses/portable-agent-setup). |

## What the glossary will not settle

It will not tell you which harness to use, whether a local model is good enough for your work, or whether a permission prompt is safe. Those are decisions about your repository, your client and your risk tolerance. The definitions only make the comparison possible.

## Related

- [Harness comparison](/harnesses) — the table this vocabulary is used to read.
- [How to test a harness in 45 minutes](/harnesses/how-to-test-a-harness) — the bake-off protocol.
- [A portable agent setup](/harnesses/portable-agent-setup) — the files that survive a switch.

## Structured data

```json
{
  "@context": "https://schema.org",
  "@type": "WebPage",
  "name": "Harness glossary",
  "url": "https://www.diginomads.win/harnesses/glossary",
  "description": "The words that decide an agent setup: model, harness, agent, tool, skill, MCP, context window, sandbox, approval policy and headless mode.",
  "isPartOf": {
    "@type": "WebSite",
    "name": "DigiNomads",
    "url": "https://www.diginomads.win/"
  }
}
```