An AI harness is everything around an AI model except the model itself. It decides what data and tools the model can reach, what it is handed with each question, what it may change, and how its answers are checked before a person acts. The model does the reasoning. The harness keeps that reasoning on your real numbers, for every chat and every teammate.
An engine on a stand can rev to the red line. It still takes you nowhere.
What makes it a car is everything bolted around it. Steering. Brakes. The gauge that says the tank is nearly empty.
An AI harness is everything around an AI model except the model itself. The model is the engine. The harness decides what it can reach, what it is handed with each question, what it may do, and how its answers are checked before a person acts.
For business data, that split often matters more than which model you pick. Put a model on a raw export and it tells a confident story. Put the same model inside a good harness and it answers from your definitions, and it says when it is not sure.
Where the word “harness” comes from
There is no standards body behind the word. Engineers have long used “test harness” for the code that runs a program and records what happened. In AI, you will see it used three ways as of September 2026.
- Evaluation harness. Software that runs a model through a fixed set of test questions and scores it. The best-known open one is EleutherAI’s LM Evaluation Harness.
- Agent harness. The loop around a model that lets it use tools and finish a task. In a November 2025 post, Anthropic calls its Claude Agent SDK a general-purpose agent harness. OpenAI describes its open-source Codex harness as what helps a model gather context, use tools, stay inside set limits and ask for approval.
- Harness design. The work of building and tuning that loop. A March 2026 Anthropic post on harness design makes a useful point. Every part of a harness rests on a guess about what the model cannot do alone. Those guesses go stale as models improve, so they need re-testing.
Most of that writing is about AI that writes code. We use the word in the same agent sense, applied to business data. The ontology (a written-down model of the business) holds the meaning. The harness governs the AI, keeping it on the right brand, definitions and numbers.
The five parts of a business AI harness
The sections below take each part, with an example from a made-up hand soap brand.
| Part | What it does | What breaks without it |
|---|---|---|
| Access | Decides which data and tools the AI can reach, locked to one brand | The AI reads a sister brand’s tables, or every table the login can see |
| Context | Hands over definitions, labels and approved facts with each question | “Revenue” means one thing in July and another in August |
| Permissions | Decides what the AI may change and what waits for a person | An ad budget moves because a chat said it should |
| Checks | Re-derives key figures, flags disagreements, refuses impossible numbers | A wrong number reads exactly like a right one |
| Memory | Writes back approved facts and decisions, never old metrics | Last spring’s return on ad spend is served as today’s |
The model sits in the middle, not at the edges. Everything before it decides what it sees, everything after it decides what a person trusts, and memory feeds the next question.
1. Access: which data and tools, locked to one brand
Access is the list of what the AI can touch: the data sources and the tools that read them. The rule that matters most is that it is locked to one brand.
Say an agency runs a hand soap brand and a body lotion brand from one login. A tool that can query every table will, sooner or later, add the lotion brand’s refunds to the soap brand’s month. Nobody notices. A good harness makes the other brand’s data unreachable, not just unlikely.
Access also means reading through fixed tools, not letting the model write any query it likes. A tool called “net revenue by month” returns the same definition every time. A model writing its own query picks a new one each time.
2. Context: definitions, labels and approved facts
Context is what the AI is handed with each question. It has three layers.
- Definitions. “Net revenue is gross sales minus discounts and refunds, by order date, before tax.” Written once, used by every answer. Our guide to AI-ready data covers how to write them.
- Labels on every number. Each figure arrives with its period, its basis and its caveats. The basis says which kind of number it is: orders placed, cash Amazon paid out, or store gross sales. A caveat might say “September is still in flight”, meaning not finished.
- Approved facts. Things no system knows on its own. The refill pouch costs $2.60 a unit to land, from June 1. The founder’s target margin is 35%. Meta ads for the soap point shoppers to Amazon, not the store.
The key design choice is that the label travels with the number. It is part of what the tool returns, not something a person has to remember to type into the chat.
3. Permissions: what the AI may change
Most business AI should read freely and change very little. The ontology lists which actions exist and who must approve them. The harness makes sure the AI cannot skip that step.
For our soap brand, a sensible split: reading sales, ads and stock needs no approval. Proposing a new unit cost is allowed, but it only takes effect after finance approves it. Changing an ad budget, a price or a listing is never done by the AI alone.
The Model Context Protocol tools spec says there should always be a person who can deny a tool call. The Claude Agent SDK controls tool use with permission modes and allow and deny rules. A business harness applies the idea to money.
4. Checks: how answers are tested before a person acts
Checks run after the model drafts an answer and before anyone acts on it. Three matter most.
- Re-derive figures from the source of record. The source of record is the system that actually recorded the money. For Amazon revenue that is the settlement report. If the draft says August net revenue was $52,000, the harness works that figure out again from the source data. If the two differ, the answer says so.
- Flag disagreements. Never average them. Meta says its ads drove $9,000 of sales. Amazon’s attribution report says $6,500. The answer shows both and says they disagree. It does not report $7,750, a number no system ever recorded.
- Refuse implausible numbers. A 140% refund rate or a 60% store conversion rate is almost always a join error. A report should refuse to print it until a person looks.
One more lesson from Anthropic’s harness design post. Agents asked to grade their own work tend to praise it. Separating the part that does the work from the part that judges it helped. For business answers, the check should compare against data, not ask the model whether it feels right.
5. Memory: what gets written back
Approved facts live in the ontology. The harness decides what from a chat may be written there.
Keep approved facts with a date, like a unit cost or a launch date. Keep decisions and the reason for them: “Moved $3,000 of ad budget to the refill pouch on September 2, because its margin is higher.” Keep forecasts with a number and a date, so they can be checked later.
Never store a metric as a fact. “Return on ad spend is 4.2” is true for one period. Stored as a fact, it gets served back in December beside a fresh 2.8, and the AI has to guess which one is real. Metrics such as lifetime value (what a customer spends with you over time) should be calculated fresh every time.
One question, with and without a harness
On September 3, the soap brand asks: “Did the refill pouch do better in August than July? Should we move ad budget to it?” All numbers are illustrative.
| Step | Without a harness | With a harness |
|---|---|---|
| Data it reaches | Every table the login sees, including the lotion brand | The soap brand only |
| July revenue | $38,000 (what Amazon paid out) | $50,000 (orders basis) |
| August revenue | $52,000 (orders basis) | $52,000 (orders basis), labeled “August settlement not final yet” |
| Change it reports | Up 37% | Up 4% |
| Unit cost used | $2.10, typed into a chat in March | $2.60, approved, from June 1 |
| Return on ad spend | 4.2, stored in spring | 2.8, calculated today |
| Ad attribution | Averages Meta and Amazon | Shows both, flags the gap |
| Budget change | Drafted and applied | Proposed, waiting for approval |
The model is the same in both columns. The left one compared cash paid with orders placed, used a stale cost and a stale metric, then acted. That is how a small real gain becomes a big budget move. More traps like these in why AI gets your ecommerce numbers wrong.
Harness vs prompt: one chat or every chat
A longer prompt fixes one chat. A harness fixes every chat, every teammate and every tool.
A prompt is text you paste in: “Use net revenue by order date. Unit cost is $2.60.” It helps, for that chat. But your ops lead opens a new chat tomorrow without it. In a long session, the model can also lose track of an instruction from far back.
A harness lives outside the chat, in the tools the AI calls. So the same rules apply whether the question comes from Claude, ChatGPT or a new hire. Prompts are fine for tone and format. They are the wrong place for the rules about money.
Where it goes wrong
A harness can look complete and still fail. These are the traps we see most.
Checks that can only pass
A check that confirms a login token “exists” says nothing about whether the token works. A check that compares a figure to itself always agrees. Before you trust any check, name the result that would prove it wrong, and make sure the check can produce it. An empty result is not a pass either. Zero rows can mean nothing happened, or that the table is missing.
Labels the model is allowed to drop
The tool returns “$52,000, orders basis, August settlement not final”. The final answer says “$52,000 in August”. The label existed and the reader never saw it. Treat a dropped caveat as a failed answer. Test for it: ask about a period you know is still in flight and read the reply.
Approvals nobody reviews
A queue of 400 proposals is not a control. It is a delay. People either approve in bulk without reading or stop approving at all. Keep the list of actions that need a person short, and make each one quick to judge.
How to evaluate an AI harness
You can test a harness in an afternoon, using questions you already know the answer to.
- Ask for last month’s net revenue. You know this from your books. Check the number, and check that the answer names its basis and period without being asked.
- Ask how this month is going, mid-month. The right answer says the month is still in flight. It should not compare half of September with all of August.
- Ask which product makes you the most money. Pick a brand where the best seller by revenue loses money after fees, ads and unit cost. A naive AI names the best seller. A harnessed one names a different product and shows the margin.
- Ask what your Meta ads sold on Amazon. The right answer says Meta’s pixel (its tracking code on your site) cannot see Amazon checkouts, and uses Amazon’s own attribution data instead.
- Ask it to change a unit cost. The right answer proposes the change and waits for approval. If it just changes it, the permissions are wrong.
Then run three more. Ask about a second brand from the first brand’s login, and expect a refusal. Plant a disagreement between two sources and see whether it is flagged or averaged. Finally, ask the same question next week in a different AI app and compare.
If you are comparing tools, our guide to MCP servers for Amazon and Shopify brands runs a similar test across three kinds of connection.
So pick a question whose answer you already know, and ask it today. A confident reply that names none of its basis, period or blind spots is an engine on a stand.
Synthesis is ontology software and an AI harness for consumer brands. It serves each brand’s joined data to Claude and other AI tools through MCP, locked to the signed-in brand. Every number carries its revenue basis, period and caveats, and approved business facts such as unit cost are attached to answers. Several governed actions need a person to approve them. Deep research runs check that each figure re-derives from the canonical number before the answer ships. Figures for the same metric and period that disagree by more than 5% are flagged, and reports refuse an implausible figure without an override. Metrics like lifetime value and return on ad spend are never stored as facts. They are computed fresh.
Questions people ask
What is an AI harness?
It is the layer around an AI model. It controls what the model can reach, the context it gets with each question, what it may do, and how answers are checked. The model is the engine. The harness is what makes its output safe to act on.
What is the difference between an agent harness and an AI model?
The model reads text and writes a reply. The agent harness runs the loop around it. It calls tools, manages context, asks for approval and returns the result. The same model can behave very differently in two harnesses.
Is an AI harness the same as a prompt?
No. A prompt is text in one chat, and the model can drop parts of it. A harness sits outside the chat. So the same data limits, labels, approvals and checks apply to every chat, every teammate and every AI tool.
What is an evaluation harness?
It is software that runs a model through a fixed set of test questions and scores the answers. The best-known open one is EleutherAI’s LM Evaluation Harness. For a brand, the useful version is a small set of questions whose answers you already know.
How do I make AI answers about my business data reliable?
Lock the AI to one brand’s data. Hand it written definitions and labels with every number. Make any change wait for a person. Check key figures against the source of record. Store approved facts, never old metrics. Then test it with questions you already know the answers to.