Real AI harnesses already exist, and you probably use a few. Examples: Claude Code (Anthropic), Cursor, Devin (Cognition, "the first AI software engineer", March 2024), LangChain's deepagents-cli, and capivaOS. A harness can even be engineered on top of another: capivaOS runs on top of Claude Code.
You already use a harness
You probably already use a harness and don't call it that. Claude Code describes itself as an agentic coding tool that lives in your terminal and understands your codebase (github.com/anthropics/claude-code). Cursor presents itself as your coding agent for building ambitious software (cursor.com). Both are harnesses: they wrap the model with tools, file access, rules, and execution. The model is the same one you'd call from an API. What makes Claude Code work in a real repository is everything around it.
From assistant to autonomous engineer
There's a spectrum. At the autonomous end is Devin, which Cognition launched in March 2024 calling it the first AI software engineer (cognition.com): it runs long-horizon tasks with less supervision. LangChain shipped deepagents-cli and describes it, in these words, as the batteries-included agent harness (github.com/langchain-ai/deepagents). So the word harness is already in the name the product gives itself. And OpenAI has a page titled Harness Engineering (by Lopopolo, openai.com/index/harness-engineering): the discipline has a name inside a frontier lab. It isn't one person's jargon.
Five real harnesses, side by side
| Harness | Type | What makes it a harness |
|---|---|---|
| Claude Code | Terminal coding agent | Wraps the model with codebase access, tools, and execution; runs in your terminal |
| Cursor | Editor (IDE) coding agent | Brings the agent into the editor, with project context and file editing |
| Devin | Autonomous software engineer | Runs long-horizon tasks with little supervision |
| deepagents-cli | Agent harness (LangChain) | Ships with the harness pieces included; its own description uses the word harness |
| capivaOS | A harness on a harness | State-machine enforced, spec-first, test-enforced pipeline, installed as a Claude Code plugin |
A harness can run on top of another
The most interesting example on the list is capivaOS, because it isn't a harness next to Claude Code. It runs on top of it. capivaOS is a state-machine enforced, spec-first, test-enforced development pipeline for Claude Code, installed as a plugin (github.com/iB2/capivaOS). Claude Code is already the harness that supplies the model, the tools, and execution. capivaOS adds the discipline layer above it: phases, gates, and artifacts. That's what harness engineering enables. You don't have to build from scratch. You can engineer a layer on a harness that already works.
Seeing examples is easy. Building yours is the point
Seeing real harnesses makes the concept concrete, but what matters is which one you build. For the coding layer you don't start from zero: capivaOS is an open-source harness on top of Claude Code, and it's a starting point. How it was engineered, phase by phase, is in the account AI Harness in Practice.
Frequently asked questions
What are examples of an AI harness?▼
Claude Code (Anthropic), Cursor, Devin (Cognition), LangChain's deepagents-cli, and capivaOS. They range from the terminal or editor coding agent to the autonomous software engineer. All of them wrap the model with tools, rules, and execution, which is what defines a harness.
Is Claude Code a harness?▼
Yes. It describes itself as an agentic coding tool that lives in your terminal and understands your codebase (github.com/anthropics/claude-code). The model is one piece; what makes it work in a real repository is the layer around it: file access, tools, permissions, and execution. That's the harness.
Is Cursor also a harness?▼
Yes. Cursor presents itself as your coding agent for building ambitious software (cursor.com). The difference from Claude Code is the form factor: Cursor lives inside the editor, Claude Code in the terminal. Both wrap the model with project context, file editing, and execution.
What is Devin?▼
Devin is an agent Cognition launched in March 2024, presented as the first AI software engineer (cognition.com). It sits at the autonomous end of the spectrum: it runs long-horizon tasks with less supervision than an editor agent. It's a harness because the value is in what wraps the model, not just the model.
Can a harness run on top of another harness?▼
Yes. capivaOS is a state-machine enforced, spec-first, test-enforced development pipeline for Claude Code, installed as a plugin (github.com/iB2/capivaOS). Claude Code is already the base harness; capivaOS adds phases, gates, and artifacts on top. It's exactly what the discipline of harness engineering enables.
Does OpenAI use the term harness engineering?▼
OpenAI has a page titled Harness Engineering, by Lopopolo (openai.com/index/harness-engineering). The name exists inside a frontier lab, which shows it isn't one person's jargon. On the origin of the term as a discipline, the one who named it was Mitchell Hashimoto, in February 2026.
You already use a harness. The next step is to engineer your own, instead of just using someone else's.