A 3D explainer for curious minds
What is a harness?
The Cambridge Dictionary says a harness is “a piece of equipment with straps and belts, used to control or hold in place a person, animal, or object.” As a verb: “to control something, usually in order to use its power.”
That orange thing floating by the climbing wall? That’s a climbing harness. In the next few minutes it will explain one of the most important ideas in AI — the agent harness.
Scroll the story ↓ — and drag the 3D scene any time to look around.
First, the real thing
The climbing harness
Climbers strap in for one big reason: safety. The harness connects you to ropes and carabiners that catch your falls, slow you down, and guide which way you go.
Its gear loops carry tools you clip on — a chalk bag, nut tools, quickdraws. And it’s yours: it travels from mountain to mountain, adapts to new terrain, and every climber, acrobat, and tree-surgeon rigs theirs a little differently.
Try it:
…or grab the rope and tools right in the 3D scene.
Now the bridge to software
Agent = Model + Harness
An AI model is like a brilliant brain with no body — it can think, but it can’t do anything on its own.
An agent harness is a piece of software that provides an environment for an AI model to operate within. Brain + harness = an agent that can actually get things done.
Look at the dashed lines: the straps that hold become rules that guide. The rope that catches and paces becomes a loop that checks the work. The gear loops that carry tools become tool plugs the model can call. Same idea — new world.
Job 1 of 4
The system prompt: the instruction sheet
Every AI model has rules baked in during training — Claude Opus 4.5 even had a famous “soul document.” A system prompt is different: it’s like the instruction sheet a new employee gets on day one. Not memorized forever — just followed while doing the work.
Watch the mail stream: the harness staples the same instruction sheet to every single message before the model reads it. That way the model always knows the rules of this job.
Try it:
Job 2 of 4
Tools: gear on the gear loops
Tools are capabilities written in code that the model can call: a web search tool, a tool that writes and runs code, a tool that composes an email.
Here’s the clever part: the harness hangs each tool on the rack with a clear label… and then steps back. It does not say when or how to use them. The model looks at the job and decides for itself.
Try it — tap a tool (here or in the scene):
Wait a moment and the model will pick one all by itself.
Job 3 of 4
The agentic loop: watch an agent work
Imagine a harness that works over email. You ask: “Compare the rankings and test scores of local primary schools, and recommend one.” Then the loop starts spinning…
- 📩 The ask arrives
- 🧠 Understand it
- 🔎 WebSearch
- 🤔 Not enough? Loop back!
- 🧮 WriteCode builds a spreadsheet
- ⚖️ Compare with the ask
- ✍️ ComposeEmail (+ attachment)
- ✅ Check own work → done!
Press run and watch the glowing dot think.
The magic is step 4: nobody told it to search again — it checked its own work and decided to. That decision to go around again is the “agentic loop”.
Job 4 of 4
The translation layer: any model fits
A harness doesn’t have to marry one AI model. The translation layer is a socket: unplug one model, plug in another — a model from Anthropic, from OpenAI, or an open-weight model that’s cheap per task — and the harness just keeps working.
Why it matters: the power moves to you. Send the same school-comparison email to three models, compare the answers and the cost, and keep everything in one place — instead of three answers scattered across three apps.
Try it — plug in a different brain (or drag a cartridge into the socket):
Same email task — three answers, one place, you pick the winner.
The part that matters most
You can own your harness
You can’t own an AI model — but you can own your harness. Like a climbing harness, you make it your own.
Pi is a minimal harness: short system prompt, tiny tool set, out of your way. It’s free, open source, and lives on your own laptop — and its users have shared more than 5,000 extensions, molding it to their work.
The first popular harness, Claude Code, was built for coding with Claude models — not as a neutral layer. Since then, free and open harnesses like OpenClaw, OpenCode, Hermes and Pi have grown up to work with any model.
What you can now explain:
- A harness holds you safely and carries your tools — in climbing and in AI.
- Every agent harness does 4 jobs: instructions (system prompt), tools, the agentic loop, and the translation layer.
- Owning your harness keeps the power in your hands.
“We wield the hammer; the hammer does not wield us.”
Based faithfully on “What is a Harness?” (earendil.com). Built as a fully offline 3D explainer — no network needed.