Model, prompt, harness, context: the four layers of an AI agent, and the only one you cannot buy
The model and the harness of a business AI agent can now be bought. The context cannot, because it is your business. That is how Alex Schultz, now Meta's Chief Data Officer, framed it at Meta's Conversations event in London in June 2026. His stack: "think model, think prompt, think harness, and then think context." This page explains each layer in plain terms, what belongs in the context layer of an agent that sells on WhatsApp, and how Learnmind builds it.
The short answer
The model is the reasoning engine. Schultz's view is that current models have "crossed the threshold" and are already good enough for what the businesses on stage described.
The prompt is the job description. Role, tone, rules, and what the agent must refuse.
The harness is the machinery around the model. Memory, tools, the connection to WhatsApp, and the handover to a person. "Harnesses you can get off the shelf."
The context is what your business knows. Prices, stock, policies, the diary, and this customer's history. In Schultz's words: "The context layer, though, really is down to you."
In our experience, most disappointing agents have a context problem, not a model problem. They answer confidently from facts nobody gave them.
What each layer does, and who supplies it
The four layers stack. Each one depends on the layer below it. A weakness anywhere shows up in the conversation as the same symptom: a customer who gets a wrong or useless answer.
| Layer | What it is | Who supplies it | What goes wrong when it is weak |
|---|---|---|---|
| Model | The AI that reads the message and writes the reply | An AI lab, reached through the harness | Rarely the bottleneck now in customer conversations |
| Prompt | Instructions: role, voice, rules, limits | Whoever builds the agent | Wrong tone, overpromising, no sense of when to stop |
| Harness | Memory, tools, the messaging channel, escalation | Off the shelf, or built | The agent can talk but cannot act: no booking, no stock check |
| Context | Your business facts, in a form a machine can use | Only you | Confident wrong answers: old prices, invented availability |
Why the model stopped being the product
For a long stretch the industry line was that "the model is the product". Schultz told the London audience there is now "more to it". The models are very good, they will keep improving, and for the customer journeys shown on stage that day they are already good enough.
Meta's July 2026 report quotes him taking a line that, as he said on stage, "lots of people are saying", and adding his own point:
Alex Schultz, Chief Data Officer, Meta
"Agents are the dumbest they are ever going to be. If you are building for where agents are today, you are building for the past six months from now."
Meta, Beyond chatbots: The agentic economy is here, July 2026The practical consequence for a business is simple. Do not spend effort on the part that improves by itself every few months. Spend it on the part that only you can improve.
What a semantic layer is, in business terms
Schultz runs data science at Meta, and he described how his teams get AI to answer analytics questions more accurately. They keep "semantic models that explain every data set, table, and dashboard" in the company. When the AI pulls that description in before it runs a query, he said, "it gets the answers right way more than when it doesn't."
The description is not tied to one tool. Any harness can read it. The business owns the meaning of its data, and every agent borrows it.
For a business that sells through WhatsApp, a semantic layer is less grand than it sounds. It is a plain, structured account of what you sell and how it works, written so a machine reads it the same way every time. In Schultz's phrase, you have to "think about how you make it computer understandable."
| What most businesses have | What an agent can actually use |
|---|---|
| Prices on request, most treatments from about AED 500 | One entry per treatment: name, price or range in AED, duration, who it is not suitable for |
| A PDF brochure from last year | A current list with a "last checked" date the agent can read |
| Ask the manager about deposits | A written deposit and cancellation rule, with the exceptions spelled out |
| Stock in a spreadsheet someone updates on Mondays | A live feed the agent queries at the moment a customer asks |
What goes into the context layer of a WhatsApp agent
These are the eight things a WhatsApp agent needs to know. Each has an owner and a rhythm, because context that nobody maintains decays faster than any model improves. On a Learnmind build, pricing logic, escalation rules and the booking calendar are settled with you before launch, not left for the agent to improvise.
| Context | Example | Changes | Owner |
|---|---|---|---|
| The offer and prices | Services or products, price ranges, packages, what is included | Monthly | Owner or finance |
| Availability | The live diary, stock levels, delivery slots | By the minute | A system connection, never a document |
| Policies | Deposits, cancellations, returns, warranty | Quarterly | Owner |
| Qualification rules | What makes an enquiry worth a salesperson's time | When the pitch changes | Sales lead |
| Escalation rules | Complaints, refunds, medical or legal questions, anything involving money | Rarely | Owner |
| Voice | How the business speaks, in Arabic and English, and what it never says | Rarely | Owner |
| Customer history | Earlier messages, bookings and purchases for this person | Every conversation | The CRM |
| Proof | Results, reviews and case studies the agent may quote | Monthly | Marketing |
Two of these are not documents at all. Availability and customer history have to be live connections, or the agent will describe a diary that stopped being true this morning. That is why every enquiry, qualification answer and booking a Learnmind agent handles is written into a built-in AI CRM. The history is there the next time that customer writes.
Too little context fails one way, too much fails another
Too little context makes an agent guess. It fills the gaps with plausible answers. That is worse than silence, because the customer believes them.
Too much context fails differently. Schultz made the point in passing: load a coding agent with 14,000 skills and "it just collapses". More material does not make an agent smarter. It makes the right fact harder to find.
We have seen a version of this ourselves. This year we told a prospect that the amount of content they had was too much for a decent AI to work with properly. The answer is not a bigger model. It is deciding what the agent needs in order to answer customers, and leaving the rest out.
- Include what customers ask about. Prices, availability, location, policies, what happens next.
- Exclude what they never ask about. Internal procedures, old campaigns, staff notes.
- Keep one version of each fact. Two price lists are worse than none.
Every team feeds the context, and one place holds it
Tanya Cordrey, Chief Product Officer of the UK car marketplace Motorway, described the problem her team is working through. "Every part of the organization has a role to play in supplying, maintaining, improving the context," she said, and it has to stay consistent.
A large company solves that with process. A clinic, a showroom or an agency solves it by naming one person and one place.
- One ownerSomebody is accountable for what the agent knows. Usually the founder or the operations lead.
- One sourcePrices, policies and rules live in one maintained place that the agent reads. Nobody edits the agent's instructions to change a price.
- Same-day updatesWhen the offer changes, the agent changes that day. For Alcaz Media, a UAE performance agency we work with, a new case study or a pricing change is live in its agent the same day.
- A weekly read of the edgesRead the conversations where the agent declined, escalated or hesitated. Missing context shows up there first.
Your WhatsApp catalogue is context too
Schultz pointed out that a business already using WhatsApp and Meta's ad tools has handed over some structured context. An uploaded catalogue tells the platform "what's in stock, what the prices are". If you use Meta's own harness, he said, make sure the important data is uploaded. Whichever harness you use, build "a semantic layer in your company of the data and context that needs to flow into the agents for them to be smart."
Treat the WhatsApp catalogue as part of the context layer, not a separate marketing chore. If the catalogue and the agent disagree about a price, the customer sees both, and trusts neither.
Five signs your agent has a context problem
- It quotes a price you stopped chargingThe price lives somewhere nobody updates.
- It says "let me check" and never doesThe harness has no connection to the system that holds the answer.
- It answers what it should hand overEscalation rules are missing, so it improvises on complaints or refunds.
- It asks returning customers for details you already holdCustomer history is not reaching the conversation.
- Two customers get two answers to the same questionThere are two versions of the fact, and the agent found both.
In every one of these cases the instinct is to rewrite the prompt. The fix is almost always upstream, in the source the agent reads or the connection to it.
How the four layers look in a Learnmind agent
- Model: we choose it. You should never have to care which one it is.
- Prompt: written in your voice, in Arabic and English, with the refusals and handovers spelled out.
- Harness: your own verified number on the official WhatsApp Business Platform, connected to your diary and CRM, with escalation to a named person on your team.
- Context: built with you before launch and kept current after it. This is the part that makes the agent yours.
Most Learnmind agents are live within two weeks. Of everything in that fortnight, the context is the part worth taking time over.
Want an agent that actually knows your business?
We build WhatsApp agents on your own verified number, grounded in your prices, policies and diary, with escalation to a named person on your team.
Terms worth being precise about
- Model
- The AI system that interprets a message and generates the reply.
- Prompt
- The standing instructions that tell the model who it is, how to speak and what it must not do.
- Harness
- The software around the model: memory, tools, channel connections and handover to people.
- Context layer
- The business facts and customer history the agent can read at the moment it replies.
- Semantic layer
- A structured description of what your data means, so any AI reads it the same way.
Frequently asked questions
What is an AI agent harness?
The software around the model that lets it remember, use tools and reach a channel such as WhatsApp.
It decides what the agent can do. It does not decide what the agent knows.
What is the context layer of an AI agent?
Everything the agent knows about your business and your customer at the moment it replies: prices,
availability, policies, rules and history, structured so a machine reads them consistently.
Will a better model fix a bad agent?
Rarely. If the agent is wrong about your business, a smarter model will be wrong more fluently. Fix
the context first.
How much information should an AI agent be given?
Enough to answer what customers actually ask, and no more. Overloading an agent makes it worse, not
better.
Who should own an agent's context?
One named person, working from one maintained source. Every team can contribute. Only one place
should hold it.
Does my WhatsApp catalogue count as context?
Yes. Keep it accurate and consistent with what the agent says, because customers can see both.
How we checked this, and what we could not settle
Checked: quotes from Alex Schultz and Tanya Cordrey come from the transcript of Meta's session "The AI Shift: The New Era of Customer Engagement". It was recorded at Conversations in London in June 2026 and published on demand by Meta. The boxed quote is the wording in Meta's July 2026 report, and Schultz's title is the one that report gives him. The Alcaz Media detail comes from our own published case study.
Not settled: Schultz's claim that a semantic layer gets answers right "way more" often is about Meta's internal analytics, not customer conversations, and he gave no figure. We have not run a controlled test of how much accuracy a structured context layer adds for a small business. Our view comes from building agents and reading their conversations, and should be read as that.
Sources
- PlatformMeta, AI Agents and the Future of Customer Engagement, the four on-demand Conversations 2026 sessions, including "The AI Shift: The New Era of Customer Engagement"
- PlatformMeta, Beyond chatbots: The agentic economy is here, July 2026
- PlatformMeta, Conversations 2026 announcements, London, 3 June 2026
- LearnmindLearnmind results, the Alcaz Media case study
Written by Edmund Gay, Learnmind.ai, Dubai. This page draws on public sessions and a report published by Meta; quotes from Meta's speakers and report are theirs; the interpretation and the checklist are ours.