Never Work With Children, Animals or AI Agents Without a Harness

What’s the best thing about generative AI? You don’t know what it’s going to say.

What’s the worst thing about generative AI? You don’t know what it’s going to say.

That uncertainty is often the point. The real world is not neat and tidy, and marketing certainly isn’t. Strategy, audiences and creative work require judgement, context, instinct and a bit of chaos. An AI that can cope with that is a gift.

You can’t ask a spreadsheet to judge your search copy or marketing strategy. You may as well show a dog a card trick. It can register that something happened, but it cannot tell you whether it was any good.

Machine learning is brilliant at a different kind of work. It can spot patterns, rank options and predict what is likely to happen next. Deterministic code can calculate a number, apply a rule and return the same answer at 9am and 4pm. Generative AI can look at a messy brief, interpret it and make a judgement. It can find the tension, challenge an assumption or explain why a perfectly sensible idea still feels flat.

These are different engines for different jobs. Generative AI is useful when ambiguity and judgement are part of the task. Machine learning is useful when prediction is the task. Deterministic code is useful when consistency is the task.

An agent describes the job it performs, not the technology underneath it.

That matters because we have just put a new product into beta. It uses several agents to complete a larger piece of work. One might interpret the user’s request. Another retrieves the right data. Another ranks a set of options. Another performs a calculation or applies a fixed rule. Some need room to think. Some should never improvise at all.

The clever bit is not putting generative AI everywhere. It is deciding where surprise creates value and where it creates risk, then getting all those parts to work together.

That is the challenge of multi-agent orchestration. With one generative agent, an unexpected answer may be useful. At the very least, you can see what it did. With several agents passing work between one another, every decision becomes the next agent’s starting point. A small wobble at the beginning can alter a retrieval, which changes a ranking, which changes the final recommendation.

Each agent may look perfectly sensible on its own while the whole system heads in the wrong direction.

In film, they say never work with children or animals. I’d add a third: chains of AI agents (without a harness).

A multi-agent product is a bit like getting a team of dogs to pull a sledge. If you can get the animals to work together it's very powerful. Each animal adjusting to the terrain and adding a huge amount of power to the operation. However try getting 6 golden retrievers to pull a cart around a part and you can imagine the chaos. The unpredictability of each dog/agent adds power but also chaos to the system. Back to agents, you might be expecting one to output a number but it responds with a question - that throws off each agent down the line. Most challenged are rather more subtle than that but to get a reliable outcome you need a harness to make sure each agent is pulling in the same direction, clear of its rules and goals. This is what we call a harness.

In practical terms, a harness is a series of rules, tests and evaluations wrapped around the agents and the product they create. Some are deterministic: do the numbers add up, is the currency correct, has the budget changed, does the source exist? Others require judgement: is this a good answer, does it address the brief, would it make sense to a planner, is the recommendation credible?

The harness runs at every handoff. It judges what one agent has produced before that output is passed to the next. It also checks the finished experience: the final answer, the calculations and what appears on screen. An individual agent can do its job correctly while the complete product still fails, so the harness has to watch both the dogs and the sledge.

This is where domain expertise becomes critical. A subject-matter expert already carries much of this harness in their head, even if they have never called it that. They know which numbers cannot move, which assumptions are dangerous, which claims need evidence and what a good answer looks like in the real world.Building the harness means turning that professional judgement into rules and tests the system can apply repeatedly.

Take a budget. A value such as “£1 million” may begin as language. Passed carelessly between systems, an encoding or parsing problem can turn it into “31 million”. The system may still produce a beautiful and coherent media plan. It will just be a £31 million one.

The harness needs to know that the budget is not prose. It is a hard constraint. So it converts the value into a structured number and currency, stores it deterministically and checks every later decision against it. The agents can use judgement to decide how to spend the budget. They do not get to decide what the budget is.

That distinction does not come from generic AI expertise. It comes from understanding the job. A technologist sees a value being passed between systems. A media planner knows the budget is the edge of the map.

And when something like that goes wrong once, we do not add a test for one exact phrase and congratulate ourselves. We test the whole family of possible failures: different currency symbols, commas, decimals, abbreviations, character encodings and values moving between text and numeric fields.

The bug becomes knowledge about how the system can fail.The harness also lets us run the whole product through thousands of synthetic journeys before a user ever sees them. Not vague “users”, but the planner opening it on Monday morning, the CFO who cares about one number and the sceptic in the room looking for the loose thread.It asks the questions those people would ask. It changes the wording. It removes data. It introduces contradictions. It tests the obvious route, the awkward route and the route nobody would choose unless they were actively trying to break the thing.

You are not merely testing whether every dog can run. You are testing whether the pack stays together when the terrain changes. The same harness works in the real beta. It watches the handoffs, records what each agent received and produced, and helps us see where the system began to move off course. Without that visibility, debugging a chain of agents is mostly guesswork. You see the answer at the end, but not the series of small decisions that created it.

The part I like most is what happens when a beta user finds something we missed.We do not just fix the individual bug. We ask whether the harness could have caught it. If it could, why didn’t it? If it couldn’t, we add the scenario and test the wider family of failures around it.When something goes wrong, we do not assume the model has learned. We expand the harness so the product has.Real usage strengthens the synthetic environment. The synthetic environment strengthens the real product. The product ratchets up. It doesn’t wobble.That is the outcome we are working towards: not a collection of clever agents, but an intelligent pack delivering to a goal.

The aim is not to make every generative agent completely predictable. That would remove the judgement and adaptability that made it useful in the first place. The aim is to make the system dependable around it.The dogs do not need to take exactly the same step every time. They need to stay together, respond to the terrain and get the sledge where it is meant to go.

That is what the harness is for.and done well its game changing for AI. As an added benefit it also gives your platforms super powers in terms of producing product help videos, sales collateral explainers and move. Because the exact same technology used to maintain and test the system can also be used for other things.

That's the beauty of AI. Each solution made and step forward opens up a world of possibilities.