harness

in English: agent harness

advancedspreading

A harness is the software framework around an AI model that provides tools, memory, and rules, turning the model into an independently working agent.

In plain terms

A language model by itself only writes text: it cannot open files, run commands, or send emails. The harness is the program that makes all this possible: it passes the task to the model, executes what the model requests, and returns the result to it.

The word originally meant horse tack or harness, and the analogy well illustrates the essence. The horse provides the power, but without a harness it cannot pull the cart and cannot be steered. Similarly, the model provides the capability, while the harness puts it to work and keeps it in check.

This is why the same model can behave quite differently in two products. One harness provides better tools, handles long tasks more skillfully, or asks for permission more strictly. When an agent performs well or poorly, it often depends at least as much on the harness as on the model.

How it works

It compiles what the model sees: the task, the base rules, the description of available tools, and the previous context.
The model responds: either by providing ready-made text or by indicating that it wants to use a tool, for example, opening a file or searching the web.
The harness checks whether the requested step is permitted, asks for your approval if necessary, then executes it and returns the result to the model.
This loop repeats until the task is completed or some limit (time, cost, step count) stops it.
Meanwhile, the harness keeps order in the background: it compresses overly long histories, saves notes, handles errors, and logs what happened.

An everyday example

You ask the coding agent to fix a bug in your program. Before deleting a file, a window pops up: 'Do you allow it?' The harness asks this question, not the model, and the harness also runs the command if you say yes.

Who is affected

developerThey choose or build the harness, and this determines how reliable and secure the agent is. → Code

entrepreneur, executiveWhen purchasing an AI product, compare the framework alongside the model name, as it determines what the agent can do within the company's systems.

employeeThose working with an agent actually see the harness interface, its permission requests, and its limits day in and day out. → Work & HR

A common misconception

MythAn agent's capability depends only on which AI model is inside it.

In realityThe same model can perform strikingly better or worse in a different harness. Model quality matters, but tools, memory management, and checks also count for a lot.

Limits and risks

The harness does not make the model smarter: if the model misunderstands the task, even the best framework will diligently work in the wrong direction.
An agent can cause as much harm as the permissions the harness grants it, so an overly permissive setup (such as automatically approving every command) is a real risk.
Measurement results are hard to compare because a score on a test reflects the combined performance of the model and the harness.
Harnesses change rapidly and vary by manufacturer, so what works in one may not be transferable to another.

Similar, but not the same

language model – The model is the part that generates text and decisions; the harness is the program built around it that turns these into actions.
agent-keretrendszer – A framework is a toolkit from which the developer assembles an agent; a harness is the specific, functional framework in which a given agent runs, and it can also be built on a framework.
system prompt – The system prompt is only a text instruction given to the model, one component of the harness, not the harness itself.

Related concepts

Broader

AI-agent, agentrendszer

Narrower

agentciklus, tool use, permission management, context management

On the edge

coding agent, system prompt, Model Context Protocol, orchestration, sandbox

A brief history

The term originates from software testing, where a 'test harness' has long referred to the framework that runs and monitors the program under test. Applied to AI agents, it gained traction in 2025 when coding agents appeared in large numbers, revealing that the framework around the model is a distinct engineering task.

Where things stand

At the end of 2026, the major AI companies and the open-source community are each developing their own harness, and the term is increasingly mentioned in the profession as a separate field (harness engineering). Most everyday users do not yet know the word, although they use the product itself.

Status: 2026 November

In the news

Appears in 16 Agent Age stories; the curve shows weekly occurrence in the news magazine.

GitHub Copilot Fusion mixes models to approach frontier coding performance at half the costOct 28, 2026 · Agent Age
Copilot Studio's GitHub Copilot harness becomes the Microsoft Copilot harnessOct 28, 2026 · Code
IFS launches industrial AI products for reliable execution and queryable operational dataOct 13, 2026 · Ipar
OpenAI's Agents API offers Codex harness as a service with computer use; data stored only in USSep 30, 2026 · Code
GitHub Security Lab automates fuzzing with an AI agent frameworkSep 24, 2026 · Code
OpenAI Agents API in public beta: the Codex harness for cloud agentsSep 21, 2026 · Code
Harness report: AI speeds up development, but deployment failure rate is 69%Sep 15, 2026 · Code
OpenAI opens the Codex harness to developers as a managed Agents APISep 14, 2026 · Code

Also known as

harness · agent harness · AI harness · agentkeret · agent runtime environment · agent framework · scaffolding

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.