A lot of people have asked me what a “harness” is when talking about AI agents. I realized I use the word more than I actually explain it.
The easiest way I’ve found to think about it is through something familiar, such as getting a client’s books ready for monthly review.
Imagine handing that assignment to someone new at your firm.
They need access to the books and supporting documents.
They need to know how you handled things last month, which accounts need attention, what they can resolve themselves, and what needs to come back to you.
They also need to know what “ready for review” actually means.
Giving an AI that assignment involves many of the same questions.
You can tell it, “Get this client ready for review.” But that instruction alone doesn’t give it access to the right records, a way to use the accounting software, or limits on what it can change.
The software that puts those pieces together around the AI is what people mean by the harness.
In this example, it would give the agent access to the client’s records and our instructions.
It would let the agent use the software that performs the calculations. It could allow it to prepare proposed adjustments while requiring our approval before posting them.
If a bank statement were missing, it would have a defined way to handle that. If the agent encountered something it couldn’t resolve, it would know where to stop and bring us in.
And when we returned to review, we’d need to see what it changed, what supported those changes, and what still needed a decision.
That’s why two systems using the same AI model can feel very different in practice.
The setup around the model affects how much useful work it can do and how much we have to not only the agent, but also ourselves.
The part that interests me most is how much of that setup depends on knowledge we already have.
We know which accounts tend to need another look. We know why we treated an expense a certain way last month. We know what makes a workpaper ready for review.
Some of that is documented, but a lot of it actually lives in our heads or gets explained while we’re reviewing someone’s work.
Getting an agent to work well means making those expectations explicit and checking whether it can actually meet them.
Writing the instructions is part of it. The software also has to provide the access, enforce the limits, and preserve the work for review.
For an accountant wondering where to start, I’d pick one recurring assignment and ask:
What would someone need to know, have access to, and bring back to me to do this well?
Your answer is a useful starting point for deciding how an agent should work in your practice and being explicit with your answers is crucial for the agent to do the job well.