We get asked questions about this from finance leaders and sharing our answers here.
CEO: What is an AI agent harness?
Otherworld Engine™: Picture two riders. The first is on a rodeo horse with nothing to steer it: no reins, no saddle, no training. The horse is powerful, and it goes wherever that power takes it. The rider hangs on and hopes not to fall off. The second rider is on a dressage horse with a proper harness. The horse is just as strong, but now its strength goes where the rider directs it, and it does so the same way every time.
An AI model is the horse. Underneath the tools you’ve heard of is what the industry calls a large language model, or LLM. It can read, reason and write, and its raw ability is real. But on its own, it works only with what you put in front of it. Paste in a fund’s holdings and it can discuss the risks. It can’t go and fetch today’s holdings, check today’s prices, run the numbers or confirm its own conclusion. And tomorrow, it won’t remember any of it.
An AI agent harness is everything built around the model to close that gap. It has four parts.
Tools. These are the hands. A tool is anything the model can use to act. That includes an API, a doorway that lets one piece of software ask another for something, like pulling live prices from a market data provider. It includes code execution, where the model writes a small program and the harness runs it, say to calculate how much of a portfolio sits in one sector. It includes search, for things like regulatory filings and news, and databases, where your holdings and transaction records live. And it includes the business applications your teams already use, like your portfolio system or your client reporting software. The model decides which tool to reach for. The harness is what actually uses it.
Memory. A model forgets everything the moment a conversation ends. Memory is how the harness fixes that. It keeps the prior context: what has already been said and done in this task, such as the half-finished rebalancing review from this morning. It keeps user preferences, like how a portfolio manager wants the morning note laid out. And it keeps workflow history, the record of steps already taken, so the agent can pick up where it left off instead of starting over.
Workspace. This is the desk the agent works at. It holds the files, the data, the environments where code can run safely, and the systems the agent is allowed to reach. The word “allowed” matters. A good workspace might let the agent read the holdings but not touch the trading system at all.
Guardrails. These are the limits, and they are what make the rest safe to switch on. Permissions say what the agent may do alone, such as drafting a rebalancing proposal. Policies are your rules written down where the agent has to follow them, like a fund’s investment mandate and its compliance limits. Approvals are the points where a person must say yes before anything goes through, like a trade or a message to a client. Monitoring is the ability to watch and record what happened, so you can answer “what did it do, and who allowed it?”
Take those four away and you have the rodeo ride where you hold on for your life: plenty of power and no control. The harness doesn’t make the horse smarter. It makes the horse’s strength usable and controllable.
That is why this work call it harness engineering. The model supplies the intelligence, and the harness decides whether that intelligence turns into reliable and dependable work.
COO: Why does an agent need both a model and a harness when a model could just do the work?
Otherworld Engine™: In a demo, it can look that way. In real work, it can’t, and the reason is simple. Reasoning and doing are two different jobs.
The model is the part that reasons. Models come from labs around the world, such as OpenAI’s GPT, Anthropic’s Claude, DeepSeek and Alibaba’s Qwen. Each one reads everything in front of it, including the task, what’s been done so far and what came back from the last step, and then decides what to do next. That is the whole job, and the best ones do it well.
The harness is the part that acts. When the model decides “pull today’s prices for this portfolio,” the model itself can’t do it. It can only say it. The harness is what connects to the market data, makes the request and hands the answer back, so the model can decide the next step. The model decides, and the harness carries it out.
Your own firm already works this way. In most investment firms, the person who decides a trade is not the person who executes it, and neither of them is the person who checks it settled. That’s deliberate. Each job has its own skills, its own controls and its own record. An agent works the same way, with the model on the decision side and the harness on the execution side. You get three things from keeping the two layers separate.
Reliability. The harness can check and record every action, because it’s the one performing it. A model on its own can claim the work is done without anyone being able to confirm it.
Flexibility. The model is a part you can change. A harness built well can run one lab’s model today and another’s next year, without rebuilding the tools, memory and controls around it.
Clear diagnosis. When something goes wrong, you can ask which layer failed: did it decide badly, or did it carry the decision out badly? Those problems have different fixes.
It’s isn’t simply about choosing the smartest model. Models will continue to improve. But, what stays right now is the need for something outside the model that acts, records and enforces the limits. A strong model won’t rescue a weak harness, and a strong harness won’t make a weak model wise.
CFO: If two firms use the same model, will they get the same results? Or does the way it’s set up change the outcome?
Otherworld Engine™: The setup changes the outcome, and often by a lot. Same horse, different harness, very different ride.
Think back to our two riders. Put the same dressage horse under two riders. One has a well-fitted harness and a trained arena. The other has loose straps and no boundaries. The horse’s talent is identical, but only one of them performs reliably. AI works the same way. Two firms can license the exact same model and get results that aren’t close.
Four parts of the harness do most of that work.
Memory. Does the agent keep the right information in view, and let go of the rest? A cluttered desk makes even a sharp mind worse. Remember the half-finished rebalancing review from this morning. If the harness loses it, the agent starts over.
Tool orchestration. Orchestration means deciding which tool gets used, in what order, and what happens with each result. Think of a conductor. The musicians are good, but without someone coordinating them, you get noise instead of music.
Feedback loops. After each step, does anything check the work? A good harness runs a test, inspects the number or asks the model to review its own output before moving on. A weak one lets the agent say “done” while the job is half finished.
Guardrails. These are the limits and approvals we discussed in the first answer. They stop the agent from taking an irreversible step, like sending a client message, without a check.
Here’s the part that matters for your budget. For work with many steps, a strong harness around a good, but not top-ranked model can beat a weak harness around the most advanced one.
The industry’s story is that performance comes from the model, so the way to improve results is to buy a bigger one. But, they say very little about what happens when a model meets your data, your expertise and your regulations. If you only upgrade the horse, you’ve bought a faster horse ride. Before you pay for the next model, ask what the harness around your current one is doing.
CIO: My teams use “agent,” “model” and “harness” as if they meant the same thing. What’s the difference, and does it matter?
Otherworld Engine™: It matters a great deal, and the horse makes it easy to see. There are three words and three different things.
The model is the horse. It supplies the strength, the instinct and the ability to read the ground and decide the next step. In technical terms, it works out what should come next, whether that’s words, numbers or a piece of code. The intelligence lives here.
The harness is everything that turns that strength into controlled movement: the reins, the saddle, the training and the arena with its boundaries. In the first answer, I called these the hands, the memory, the desk and the limits. It’s the part that lets the horse’s decisions become actions.
The agent is the horse in its harness, working. It’s the complete system. A horse standing in a field isn’t an agent, and neither is a pile of saddles and reins in a shed. Only when the two come together do you get something that can do a job. And it’s what you actually ride. As a user, you set the direction and the task.
There are three reasons this matters for you.
First, diagnosis. When an agent gives a wrong answer, the useful question isn’t “is the AI bad?” It’s “which part failed?” Say the agent reports a fund’s exposure to one sector incorrectly. Perhaps the horse misjudged the ground, meaning the model reasoned badly. Perhaps the arena was set up wrong, meaning the harness fed it yesterday’s prices. Perhaps nothing checked the number before it went out. Those problems have different fixes. A team that can’t tell them apart ends up swapping horses for months when the problem was the harness all along.
Second, buying. The industry uses “agent” loosely, and that’s convenient for anyone selling horses. A vendor says “our agent is powered by the most advanced model,” and what you’ve learned is how good the horse is. You haven’t learned how it’s ridden. Ask what the agent remembers, what it’s allowed to touch, who approves its actions and what gets recorded. Those answers describe the harness, and the harness decides whether the agent’s safe to put near a company mandate.
Third, ownership. The harness is where your rules, your data and your approvals live. The model can change, and a good harness lets you change it. That makes the harness the part you most need to keep control of.
COO: My teams spend hours prompt engineering. Someone told me there’s “context engineering” and now “harness engineering.” Are those just new names for the same skill?
Otherworld Engine™: They’re three layers of the same job, and each one wraps the one before it.
Prompt engineering is the instruction you give. It’s choosing your words so the model understands the task. “Summarize this fund’s risks” is a weak instruction. “Summarize this fund’s three largest risks for a client who’s worried about interest rates, in five sentences” is a strong one. This is the first and oldest layer. Think of it as the command you give the horse as the rider. A clear command beats a vague one.
Context engineering is deciding what the model gets to see, and when. A model can only hold so much in view at once, and what’s in view shapes what it concludes. This is where retrieval comes in: the harness fetches the right documents and hands them over, so you don’t have to paste everything in. Think of it as what the horse can see ahead. Show it the next three fences and it clears them. Show it the whole county and it loses the path.
Harness engineering is designing the entire setup around the model. That includes the prompts and the context, plus the tools, the safe place to run code, the checking after each step and the limits and approvals. It’s the reins, the training and the arena, all at once.
Prompts and context both live inside the harness as parts of it. Good instructions still matter, and a team that expects instructions to do the work of a whole system will be disappointed.
Over the last few years, the focus has moved outward: first how to ask, then what to show, now what to build around the model. It moved because wording alone hit a ceiling. You can word an instruction perfectly, and the model still can’t fetch today’s prices, run the calculation or check its own number. No wording can supply a capability the model doesn’t have.
There’s a business around the first layer, with prompt libraries, prompt courses and prompt specialists. That keeps your teams working on the smallest piece of the problem. A perfect command to a horse in an empty field still gets you nowhere. Before your teams spend another week on wording, ask what the agent can fetch, what it can check and what it needs approval for. This will move your results.
CEO: Can you give me a simple definition of each part of a harness?
Otherworld Engine™: Yes. They follow the workflow of a task, from the instructions at the start to the outcome at the end.
1. System prompt. These are the standing instructions the agent reads every time it starts. They say who the agent is, what it’s for and what rules it follows. For a research assistant at a fund, that could include a rule never to quote a price without naming its source. Vague instructions are one of the most common causes of an agent behaving differently from one day to the next.
2. Tools. These are the things the agent can use to act: searching, querying a database, sending an email, calling an API. The model chooses the tool, and the harness runs it and hands back the result. Many teams used to build a separate tool for every small job, such as one to fetch a price and another to fetch holdings. Now they’re increasingly giving agents one general ability instead: write a small program and run it. That lets the agent put together the exact steps a job needs, including steps nobody planned for.
3. Sandbox. This is a sealed practice room. The agent can run code and try things there with no effect on your real systems. If something goes wrong, you reset the room or shut it down. You can also run many agents at once, each in its own room. A portfolio calculation, for instance, can be tested on a copy of the data.
4. File System. This is a filing cabinet that keeps the agent’s work between sessions: code, notes, plans, half-finished analysis. People and other agents can open the same cabinet, so a draft started on Monday is still there on Tuesday, and an analyst can read it and add to it.
5. Memory and context management. A model can only hold so much in view at once, and that limit is called the context window. The harness decides what stays in view and what gets summarized and filed away. That summarizing is called context compaction. Skip it and the window fills with clutter on a long task, and the quality of the thinking drops. That decline has a name too: context rot. Across sessions, the harness stores and retrieves the relevant history, so the agent can resume where it left off.
6. Feedback loops. After each step, someone checks the work. The harness can run a test, inspect the result or ask the model to review its own output, which is called self-verification. A simple example is confirming that a portfolio’s weights add up to 100% before the report goes out. Without a loop like this, agents tend to declare the job done too early.
7. Guardrails. These are rules that block unsafe or unapproved actions. One common kind is human-in-the-loop, where a person must review and approve before an action goes through. An agent can draft a trade proposal, and the portfolio manager approves it before anything happens. In enterprise settings, these checkpoints are often mandatory.
8. Observability. This is the ability to see what the agent did, why it made each decision and where things went wrong. It runs on logs, a written record of each action, traces, the step-by-step path from request to result, and dashboards, a screen where you can watch it all. Regulated industries need an audit trail showing what the agent did and on whose authority. At scale, the same records feed evaluation: measuring whether agents perform correctly across thousands of runs of real work, which a demo can’t show you.
When agents fail in production, the cause usually sits in one of these parts. A long task fills the context window with clutter and the answers get worse. The wrong document gets retrieved, and the agent builds a confident answer on top of it. Nothing checks the work, so a rebalancing gets reported as done when it’s half finished. A client message goes out with no approval step in front of it.
Each of these gets blamed on the model. The failures live in the harness, so that’s where the fix lives.