I ride a horse that knows every aid in the book and still invents his own answers about half the time. The bridle and reins gives him boundaries. The training gives him judgement. On a good day the two add up to a horse I can trust around a course. On a bad day the harness holds while the training cracks, and I find myself airborne and in the dirt in front of the fence I was meant to go over.
I’ve been thinking about this a lot lately, because building with LLMs feels exactly the same.
The word “harness” is what the industry has settled on for the scaffolding around an AI model — the prompts, the tools, the approval gates, the bits of plumbing that decide what a model is allowed to do and how a human stays in the loop. The harness is important but it’s not what makes a good agent, any more than tack makes a good horse. The training does. The thousand small decisions about how the model has been instructed, what skills it can reach for, what counts as a good answer in this context and not another — that’s where the work is.
Right now I’m building the training and the tack. And today I’m sharing the first version of WooAgent OS — a team of AI coworkers for your WooCommerce store.
WooAgent OS was designed and built in collaboration with Nevena Ilic as part of Automattic’s Radical Speed month.
What it is
The concept is simple: agents propose, operators approve. Each agent on the team has a job. Marketing rewrites product copy, Pricing benchmarks against comparable stores, Sales Support drafts customer replies — and every piece of work lands as a row with a before/after diff: what’s there now, what it wants to change, why.
For example, the Pricing agent watches competitor prices on the web, and when it spots a real gap, it posts a proposal — “drop this throw pillow from $48 to $39” — with multiple retailer URLs as evidence and a section for reasoning. You read it. You approve, or you reject.
That’s the harness. The interesting part then is the training. How does the Pricing agent know which competitors to look at, and how does it tell a real price gap from a sale that ends tomorrow? How does it phrase a proposal so a busy operator can decide in ten seconds? That’s a lot of the prompt tuning we’ve been working on in recent weeks, and there’s a lot more of that to come.
Why now
Most WooCommerce admins already have AI in them somewhere. A chatbot for customer questions. A description generator. A suggestion engine in your SEO plugin. WooAgent OS isn’t another one of those tools — it’s the team above them. It’s a small fleet of AI coworkers that drafts work across your store and posts each change to a single queue for your one-click approval.
It’s not autonomous, not a chatbot, and not a replacement for the operator. WooAgent works for you and waits for you to say yes. It’s slower than fully automated agents on purpose so operators can have full trust.
WooAgent is open-source, local-first, and your data stays on your machine.
What’s deliberately rough
This is early. A few things I already know need polish:
- Onboarding. The install path works but it’s not what I want it to be yet.
- Persona coverage. v0.1 ships three agents — Marketing, Pricing, Sales Support. Four more (Reporting, Accounting, Inventory, and the Chief of Sraff) are sketched but not live.
- Trust signals. The proposal queue needs better evidence rendering — operators should be able to verify a claim without leaving the screen.
Please test it. Please tell me where it breaks. The harness is in your hands now; I’m still working on the training.
GitHub: github.com/Automattic/wooagent-os
Thanks for reading.