What a harness is, and how AURORA's works.

When people talk about AI agents, they almost always talk about the model: which one is smartest, which one costs less. But the model is only one part. What decides whether an agent works well is what surrounds it. That has a name: the harness.

The model alone is not enough

An AI model takes in text and gives back text. It is very capable, but on its own it doesn't know where it is. It doesn't know which files it can open, what it is allowed to change, when to stop or whom to ask.

A harness is what connects a force to useful work and keeps it safe: think of a climbing harness, or the harness on a horse. In AI it is the same. The harness is the structure around the model that turns it into an agent able to finish a task.

In an article on AI models we called it the operational scaffolding. Same idea.

What a harness does, in five questions

A good harness answers five questions before the model writes a single line:

• What does it know? The instructions and context it gets for that piece of work.
• What can it use? The tools it is allowed: reading a file, editing it, looking at a page.
• Where can it write? The permissions, and a workspace kept apart from the public one.
• When does it stop? The rules that say when to deliver and when to ask.
• How did it go? The trace of what it did, and the criterion the result is measured by.

AURORA's harness, step by step

AURORA is Tsubasa's marketing agent. Its harness is easiest to see by following one piece of work from start to finish.

1 · The request. You write to her in plain words, from the reserved area or from Telegram, in Italian, English or German. The work lands on the Agent page and moves through clear phases: received, in progress, delivery, test site, production.

2 · The context. AURORA only uses your facts: what is on the website, the documents you give her, the strategy, the brand kit. If a piece of information is missing, she asks you instead of making it up.

3 · The rules. The brand kit and the design system are not just for people: she reads them too. Colours, voice and components become rules every delivery follows. Then there is the method: the marketing rules she follows in every piece of work.

4 · The right model for each task. Inside the harness, tasks are handed to different models depending on what is needed. A simple task doesn't tie up the heaviest model; a hard task doesn't go to the lightest one.

5 · The test. AURORA prepares the change, or the proposals for a new page. She puts the delivery on a test site identical to the public one and shows you a photo of the page.

6 · The decision. Here the harness stops and hands over to you. You decide which proposal to approve, what goes to production and when, who joins the conversations, and when a piece of work is closed.

What changes for you

• The website stays consistent. Each step builds on the one before: strategy guides the brand, the brand sets the rules for the design system, the design system shapes the pages. Even after many changes.
• No invented facts. If a fact is missing, you get a question, not a filler sentence.
• No surprises online. The work reaches the test site on its own. Only what you decide goes to production.
• Work doesn't start from scratch. What you give her and what she delivers stays in the reserved area, ready for the next piece of work.

What the harness doesn't do

The harness doesn't make the model infallible. It makes its work readable and sets clear boundaries: where it can act, what it must ask, what stays with you.

That is why AURORA grows in stages. In the next ones, each flow will gain autonomy only after it has produced correct results and left a readable trace of its decisions.

An agent is only as good as the pitch it plays on. If you want to see what AURORA does today and what you decide, it's all on the AURORA page.