Why Kaz has his own harness

AI might not do everything, but large language models are very good at many tasks, including working with tools. We are the first adtech company to have launched a dedicated custom ad operations harness, and we'll explain why in this article. We call our agent Kaz, and he has already become a valuable addition to both our internal operations and how our publishers run their ad setup.

What matters about a custom harness

What matters about a custom harness is that Kaz does not browse your account like a normal person. Instead, we provide him with structured tools for exploring, configuring, querying, and analyzing reports, simulating and proposing rules and placements, and running experiments and more.

Our goal is always to keep feature parity, but not necessarily the exact same approach. We want Kaz to be able to do everything a human can do in our platform, but there are different methods for how agents interact with the space around them compared to how we do.

Other platforms may choose to use general MCP, the Model Context Protocol. This is a type of link that agents in other apps like ChatGPT or Claude can call to collect dedicated information from a platform. This is useful for pulling numbers into tools, but it doesn't give the agent a concrete description of how things should be done or validation and support for creating things and inserting them onto your page. We still support this, and it might be helpful to you if you're already analyzing data elsewhere. MCP works great as a pipe, just like any other reporting API, which we also make available to all publishers, but it can't do everything.

Performative agents may look busy

Performative agents may look busy. They can narrate. They might use a virtual computer to click around, but they don't have native tools that are tailored to how, in this context, ads should be managed on your website.

The tools don't listen to natural language

Managing ads is a high-stakes task, but not necessarily a difficult one for 2026's models. Through our harness, we can achieve maximal trust because it tells (and requires) Kaz to follow the rules.

It means things can't pass through without your approval because the tools Kaz uses don't listen to natural language. They parse things into machine-readable data and code that you always approve or have us approve for you. It also means he never touches payouts, or the rest of the account.

Ad ops is not a job for a generalist

We believe that ad operations aren't a good domain for a generalist chat app or even a coding agent. A visitor is not a row in a spreadsheet, and there are many ways to track or interpret one. In our case, we might look at factors like their geography, device, the kind of page they are on, whether they consented to cookies, and many other points.

A harness that knows exactly what we look for and how our rules engine is structured makes simulation cheap and an external chatbot hard to keep under control.

Making our own lives easier

We built this because we also like making our own lives easier. Agent demos are quite fashionable, but most agentic implementations don't really make your life easier. Putting a coding agent in charge of your computer is slower than doing things yourself, at least for now. On the other hand, building tools around what is sometimes annoying and sometimes just unnecessary or more difficult for a human is very valuable.

We keep humans in the loop, and in fact, we are happy to run Kaz for you. We don't pitch to replace your team, but we do like avoiding chores and all of the boring stuff.

If the chores are the problem, that's the kind of thing Kaz is for. Open the workbench, or ask us to run him for you — the contact page is the shortest path to either.