Harness, Loops, Evals and RAG: Build AI Agents Like You Hire

Prompting is cheap. Everything around the AI is not.

Right now, half the world is learning to write better prompts. Prompting is about the cheapest skill there is: a 15-year-old can get good at it in a few weeks. The expensive part, the part almost nobody teaches, is building everything around the AI. There are four buzzwords for it: harness, loop engineering, evals and tracing. They sound like PhD material. They are not. Each one is something you already do when you bring a new person onto your team.

This matters in Malaysia specifically. The World Bank figure I cite in the video puts the number of people in Malaysia who can genuinely build with AI at around 3,000, with demand heading towards 30,000. That gap of roughly 27,000 will not be filled by the best prompt writers. It will be filled by people who can do what this article describes.

What a model actually is, and the four holes

Strip it down. An AI model is a machine that guesses the next word. It has read almost everything humans have written, which makes its guesses extremely good, but underneath it is still guessing. Now think about what that machine does not have. Write these four down:

The harness: onboarding a genius

The word for filling all four holes is the harness, borrowed from horse riding: the horse has the power, the harness lets you point it somewhere. But the real problem is onboarding. You have just hired a genius who has read every book in the world but does not know your name, your customers, or where the toilet is. You give them exactly what you give any new hire:

First, the handbook. How we do things here, our mission, our tone of voice, what we never say to a customer, and that anything above $500 gets escalated. In AI land this is the system prompt, or increasingly a skill: a document, often a markdown file, with instructions.

Second, the files. The facts: who our customers are, what we sell, our refund policy. When the agent looks things up in those files before answering, that is RAG, retrieval-augmented generation. It just means checking the folder before replying to the customer.

Third, the log. What actually happened. This customer called twice last month, she was angry, we gave her a voucher, she is still angry. Writing it down, timestamped, gives the agent context and gives you something to verify. Handbook, files and log together are the whole memory system. For a deeper dive see our guide to context engineering.

Hands, and the most expensive instruction: "just handle it"

Next you give it hands: access to pull up the customer database, book meetings, issue refunds. Now imagine you onboard an intern and say, "Go handle the customer complaints." They find 50 complaints and start refunding everyone who is unhappy. When do they stop? At 30? 500? When do they come back and ask whether they are doing it right?

Anyone who has managed people knows that "just handle it", with no clear steps or end goal, is the most expensive instruction you can give. A good manager says: do these ten steps, and before step eleven, check with me. That is loop engineering. An agent works in a loop: think, act, look at the result, think again. Loop engineering is you deciding when the job is done, how it is done, and when the agent must stop and ask. It is the single most important part of the system, and it is literally just proper delegation. We go further in loop engineering.

Here is what it costs to get wrong. An agent with no stopping rule does not crash. It keeps working, quietly and confidently, for hours, burning hundreds of ringgit in token costs on something nobody asked for. Exactly like a very bright junior who is too scared to ask questions.

Checking the intern's work: tracing and evals

Then you check the work. That maps to two unglamorous words. Tracing means recording every step of every run: what it looked at, what it decided, which tool it used, how long it took and what it cost. Without that, you are managing someone whose work you have never looked at. You are just vibing.

Evals are a performance review rubric. You write down what a good answer looks like before you see any answers, then score every run pass or fail against it. Right now the standard test in most companies is "this version feels better". That is not a test, it is a feeling. You would not promote someone based on a feeling. You agree a scoreboard in advance and look at the numbers.

Put tracing and evals together and you get what people call LLMOps: watch what it did, score it, find the broken bit, fix the handbook, run it again. Think of it as a probation review that never ends.

The model is the cheapest part

Once you see all this, you realise the model is the cheapest part of the equation. Your competitor has the same models you do. So does a kid on the other side of the world. Everyone can rent the same brain for a few hundred ringgit a month. There is no moat in the model.

What your competitor does not have is your customer log, your handbook, your stopping rules and your scorecards for what a good answer looks like in your business. All four are management work, not machine learning research.

That flips the usual AI jobs story. Everyone fears AI taking the job. What is actually happening is that AI creates a pile of work somebody has to do: writing SOPs, setting boundaries, reading logs, verifying output, scoring results. That job already exists. It is called being a manager. It just has a new kind of report now, one that never sleeps and has no common sense. And building these systems reliably for a real business is software engineering: tools, data, permissions, logging and tests. We break the role down in what AI-native developers actually do.

Why four in ten projects fail, and your homework

So why do so many AI projects get cancelled? Not because the AI failed. Because the company hired a genius brain, sat it at a desk with no handbook, no logs, no stopping rule and nobody reviewing its work, and then acted surprised when it went wrong.

Here is your homework for this week. Pick one task you do every week. Write a one-page handbook for it. Give an agent one tool. Write down one stopping rule. Run it ten times and mark each run pass or fail. That is the entire system at small scale, and honestly it is more than most companies have ever done.

If you want to go from running that experiment to building agents that hold up in production, with real code, real data and real evals, that is what our 12-week AI-Native Software Development Programme teaches, starting from zero. You can try the first lessons in the free trial. And if this was a lot at once, start with the gentler version: AI agents explained like you're 10.