What Is an Agent Harness? The Model Left My Plan and Nothing Broke

Two weeks after it shipped, the model I’d been building with — Fable 5 — left my Claude subscription plan. Anthropic’s schedule, not mine.

I’d used it to build a 240-listing directory a few hours before the window closed. Prime conditions for a very bad week.

Instead, nothing broke. The directory stayed live, the content pipeline kept running, and I lost zero workflows. The reason has a name, and the name is the point of this post: the agent harness.

I’ve already written the wide-angle definition of what an AI harness is — one system instead of 20 tools, the five layers, the whole map. This page zooms in on one layer of that map: the harness around a single AI agent, and why it’s the part of your setup that outlives any model.

A model is a brain. The agent harness is the body.

An AI agent is a model plus a harness. That’s the whole equation.

The model is the brain — rented intelligence. The harness is the body around it: the instructions it reads before every session, the tools it’s allowed to use, the memory it keeps, the limits on what it can touch.

Under the hood, every agent runs one loop: read the situation, pick a tool, run it, check the result, ask “am I done?”, repeat. The harness is whatever keeps that loop running reliably so you’re not babysitting it.

A genius brain with no harness is still just a chatbot. A weaker brain inside a great harness will still ship. That explains most of what you’ve watched AI do well and do badly this year.

You’ll see the same idea under other names. Developers benchmarking models say LLM harness; some tutorials say AI agent harness. Same body, different camera angle.

The week the model left my plan

Back to the Fable 5 story.

This isn’t a review of that model. The relevant fact is the timeline: shipped, then scheduled out of my plan two weeks later. Models do that now — providers deprecate them, reprice them, or move them behind different tiers on their own schedule... and you don’t get a vote.

What mattered was where my work actually lived. The directory pipeline wasn’t inside the model, and it wasn’t inside any vendor’s dashboard. It was a folder of plain files: standing instructions, skills, sub-agents, a review step. Claude Code was the runtime. The model was one setting.

So when the window closed, I changed the setting. A different brain went into the same body, and the pipeline kept shipping. A community in El Salvador submitted a listing through the directory’s form a few days ago, and the harness handled it without me.

(That’s the real test of an agent harness: does it keep working when something upstream changes without asking you first.)

What’s literally inside one

If you open one of my projects, the harness is visible as a folder.

In Claude Code it’s the .claude directory plus a CLAUDE.md file — the standing instructions the agent reads at the start of every session. Next to those live the specialist agents, the skills (written procedures the AI loads when a task matches), commands, and hooks. That structure is the body.

The piece that makes it scale is sub-agents. Instead of cramming every document into one context window, an orchestrator hands each fresh sub-agent a file path and one job. The template I borrowed — Scott Graham’s safe-agentic-workflow (cheddarfox on GitHub), free to clone — runs 11 of them. My own AEO-audit rig runs 10.

Why bother? Because a stuffed context window makes the model dumber. Even with a million-token limit, loading everything at once burns hundreds of thousands of tokens, and quality drops as the window fills. Small, fresh contexts keep each worker sharp.

And the layer I’d never skip now: verification. One of my harnesses includes a reviewer agent that scores drafts against a written voice standard before anything ships. A guide scored 34/50 and got bounced back for a revision pass; a different one hit 45/50 and went straight through. That scoring step catches what I’d have missed at 11pm.

Swap the brain, keep the body

The belief underneath all of this: own your agents, swap your models.

The practical version — build with frontier models, execute with cheaper ones. Use an expensive top-tier model for the one-time design work: writing the instructions, structuring the sub-agents, wiring the review step. Then run the day-to-day jobs through that same harness on cheaper or open-source models.

The design reasoning happens once. The execution happens hundreds of times. Pay frontier prices for the first and pennies for the second.

I route mine through the Venice API so mixing models per task is a settings change, and my content harness now runs a full pass on open-source models... output that would have cost a few dollars in frontier tokens every single time it fires.

None of this works if your workflow lives inside one vendor’s chat window. All of it works when the harness is files you own.

The honest part: mine breaks. On camera.

I recorded myself building one of these from scratch, so I have the failure reel in writing.

I closed the running harness by accident at the 21–22 minute mark. (“Whoops... it’s gonna start over.” It did.)

The newsletter draft got routed to the wrong destination mid-demo, so I told the agent to fix that part of the harness so it couldn’t happen again, then re-ran it. The YouTube metadata came out wrong because I’d pointed the thing at an old transcript. A couple of skills weren’t where the harness expected them, and the agent had to go hunting.

And I still haven’t solved the clean-state problem: run harnesses in two parallel conversations and the git history turns into spaghetti. (If you’ve solved that one, genuinely, tell me.)

That’s the texture of building one. A harness is a thing you fix and re-run — the version that never misroutes a draft only exists because an earlier version did.

Where to start this week

• Look at the body you already have. Open Claude Code and inspect the .claude folder and CLAUDE.mdin any project. If they’re empty, that’s your starting point, not a failure.
• Steal a good one. Clone the Safe Agentic Workflow template and read how the orchestrator hands jobs to fresh sub-agents. Reading a working harness teaches faster than building blind.
• Add one verification step. A reviewer agent that scores output and bounces weak drafts will catch the mistakes you stopped noticing.

For scale: a full run of my content harness takes 35–37 minutes end-to-end and produces real, voice-checked drafts. Weeks of tinkering got it there, and it still isn’t finished. Honest timeframes or nothing.

In the Academy’s Armory, our Rigs are exactly this — battle-tested agent harnesses for one business job each (SEO audits, funnel builds, content repurposing) that members customize instead of starting from an empty folder. The full Armory is included at every tier of the Academy ($83/month or $497/year, 7-day free trial), along with feedback on the harness you build with it.

Questions people actually ask

What is an AI agent harness?

An AI agent harness is the working structure wrapped around an AI model: standing instructions, context files, the tools the agent may use, memory between sessions, and guardrails on what it can touch. The model supplies intelligence; the harness turns that intelligence into finished, repeatable work. It's the same concept as an agent harness — the extra 'AI' just distinguishes it from hardware test harnesses.

What is an LLM harness?

An LLM harness is the same idea named from the model's side: the scaffolding a large language model runs inside — prompt, tools, loop, and evaluation. Developers use the term when testing or benchmarking models; business builders usually say agent harness. Either way, the harness decides whether the model's raw output becomes reliable work or stays a clever chat transcript.

What's the difference between the agent harness and the model?

The model is the brain: rented intelligence from a provider, replaceable and regularly replaced. The harness is the body: your instructions, tools, memory, and checks. Without a harness, the smartest model on earth is a chat window; inside a good harness, a mid-tier model ships finished work. When a model gets deprecated, the harness is the part that survives.

Can I swap models inside the same agent harness?

Yes — that's the point of building one. When the model I'd been using left my subscription plan two weeks after it shipped, I plugged a different model into the same harness and every workflow kept running. If your harness is plain files you own, changing models is closer to a settings change than a rebuild.

Do I need to be a developer to build an agent harness?

No. I'm not a full-on developer and mine run real production work. In Claude Code, a harness is mostly a folder of plain-language files: a CLAUDE.md with standing instructions, plus skills, agents, and commands. Cloning a proven template — Scott Graham's safe-agentic-workflow (cheddarfox on GitHub) is a good one — beats starting from an empty folder.

What does an agent harness actually contain?

In practice: standing instructions the agent reads every session, context files about your business, tool access with permissions, reusable skills for repeat processes, sub-agents that take delegated pieces of a job, and a verification step that checks output before it ships. Mine bounced a draft that scored 34/50 on a voice review and passed one at 45/50 — that checking layer is the difference between drafts you trust and drafts you babysit.