What is a harness?

The layer that allows your language model to get real work done. The thing that turns a chatbot into an operator that gets smarter every day.

It is Tuesday morning. Before I'm out of bed, a brief lands on my phone. The thing that wrote it had already read overnight orders, cross-referenced them against my support inbox, summarized what's open, drafted three customer replies, flagged the one issue that needs me, and texted the whole package to Signal.

The model that did the writing was Claude. The thing that knew which folders to read, which inboxes to summarize, which voice to write in, and where to send the result was the harness. The model is like the engine. The harness is like the car.

When you hear people talk about agents, they’re talking about the combination of a model and a harness. LLM + Harness = Agent.

What it is

A harness is the operator-level layer around a language model. A model on its own is text in, text out (ya, sometimes images and video too, but you get the point). It is useful in a chat tab and insufficient for running parts of your business autonomously. A harness adds four things the model doesn't have on its own:

• Memory. Files the harness reads on every session so the model knows your business at hour zero. Who your customers are, what ARR means in your company, what the canonical files are, what your week looks like. Without memory, every conversation is a re-introduction.
• Tools. Connectors to your real systems. CRM, inbox, calendar, file storage, Slack, Stripe, whatever your business runs on. The harness gives the model the ability to read from those systems and write to them, not just talk about them.
• Skills. Named capabilities the model can invoke. Pre-meeting brief. Commitment ledger. Pipeline radar. Each one is a folder describing a workflow the harness knows how to run on command.
• Routines. Scheduling. The harness fires skills on a cadence (Monday brief at 7am, pipeline check daily at 6am, Friday wrap at 5pm) without you having to open it. Wakes up to you, not for you.

The clearest examples in the operator community today are [OpenCLAW][1] (the open-source harness most CEOs in the operator community run; the one I run at Headphones.com and Lantern.is) and Hermes (Nous Research's open-source alternative; mobile-first). Claude Code itself is a thin harness. OpenCLAW is a richer one built on top of the same primitives.

Why it matters

I lived in AI chat tabs for about a year before I installed a harness. I'd paste context in. I'd paste outputs out. Every conversation started cold. Every Monday was a fresh start. The model was useful and the leverage was bounded by how much copy-paste I was willing to do in a day.

The day my agent first texted me a Monday brief before I'd opened my laptop, I realized I'd been running a fundamentally different operation the year before.

The model is interchangeable. The harness is what compounds. Pretty much all CEOs have paid Claude accounts now; a great model is almost a commodity. The harness is where the leverage lives because the harness is what holds memory, calls connectors, runs skills, and wakes up on schedule. Every week of running a harness compounds. Every week of running chat tabs basically starts from scratch.

The CEOs who install the harness and learn how to optimize it will outperform chatbot users exponentially over the next few years. The compounding is invisible from the outside. By the time it's visible, the lead will be enormous (and still compounding).

What a good harness setup looks like

The four pillars in operator vernacular. None of them require a developer.

• Memory in tiers. Tier 1 is durable (who your business is, what your week looks like, who your top customers are; rarely changes). Tier 2 is quarterly (current OKRs, current customers, current priorities; refresh every 90 days). Tier 3 is session-level (what you're working on right now; automatic). The harness reads all three at startup. Tier 1 is the install everyone does; Tier 2 is where compounding lives; Tier 3 is gravy.
• Connectors only after a skill needs them. Don't wire your CRM API on day three. Wire connectors when a skill genuinely calls for them. Otherwise you've got infrastructure looking for a job and a maintenance tax with no benefit.
• Skills as the unit of work. Named workflows in folders. Pre-meeting brief, commitment ledger, pipeline hygiene, team and investor updates. Each one is a SKILL.md describing what it does and when to invoke it. Skills follow the open agentskills.io standard, so they're portable across harnesses; the skill you write today works on whichever harness you end up running.
• Routines that wake the harness up. The Monday brief that lands at 7am. The Friday wrap that runs at 5pm. The pipeline check that fires daily at 6am. Without routines, the harness is reactive (waits for you); with them, it operates a calendar of its own.

A working harness looks like infrastructure, not a product. By month two it stops being a project you work on and starts being the system your week runs through.

Common mistakes

• Installing without committing to skills. A harness without skills is a folder of disappointment. The harness only earns its keep when the skills are doing real work. The infrastructure isn't the product; the workflows are.
• Wiring connectors too early. New harness owners often spend the first weekend connecting every API in their stack. Most of those connections sit unused. Wire connectors when a skill needs them; not before.
• Treating it like a SaaS. A harness is not a product you sign up for. It is infrastructure you build into your operating cadence. It takes weeks to get right, not minutes. CEOs who expect SaaS-grade onboarding will be disappointed.
• Skipping memory refresh. Tier 2 drift is the silent killer. Quarterly refresh of OKRs, customers, priorities is the discipline that earns the harness its compounding. Skip the refresh and the harness starts making decisions on stale data.
• Expecting it to replace your judgment. A harness is leverage on your judgment, not a substitute. The agent drafts the brief and YOU decide what to do with it.

Do this next

If you want to see what a working harness looks like in practice, start with [Granola → markdown][2]. It is the foundational pipeline every other workflow in this stack reads from. It should take about 30 minutes. If you'd rather start with the harness itself, [What is OpenCLAW?][3] gives you an introduction to one of the most popular harnesses these days.

[1]: /workflows/what-is-openclaw [2]: /workflows/granola-to-markdown [3]: /workflows/what-is-openclaw