Strapping a Formula 1 engine onto your old wheelbarrow to make it go will probably not end well, but not because the Formula 1 engine is not top-notch.
Okay, okay, let me back the wheelbarrow up a bit.
At FRG, we are thinking about and figuring out how to get the most out of AI when it comes to measuring and managing risk. That makes for a lot of reading, conversations, and proving and improving concepts.
Last week, Tim Weeks, FRG’s managing partner, sent me an article distilling the latest and greatest on AI Agent Harness Engineering. While that might not sound like most people’s favourite bedtime story, it makes for some solid points.
Just to get the terminology straight:
An AI agent comprises the parts/models that perform the actual task.
For example, it can keep an eye on when new data lands, then call the part that pulls the data in, monitor and remedy any data issues, send an alert if the issues signify something systemically off, call the part that crunches the numbers, again keep an eye on the process and the results, investigate hiccups, attempt to fix them, alert the humans that this happened, call the part that formats the results and distributes them, get and respond to feedback, and then watch for new data so it can do it all again, just better.
The harness is all the instructions, rules, and guardrails that govern how the agent does it, including:
- Instructions on the order of execution and what it should do if a step fails, and even what constitutes a failure.
- An inventory of tools/parts/models to use for each step.
- Rules for the information/context that each tool and the agent itself need to run and keep for learning.
- An audit trail that remembers every state the agent has ever experienced and what was done about it, and if it was the right thing to do, and if not what it should have done instead.
- Safety and security information for and about the agent and its parts.
- Instructions for measuring the agent's performance. Is it running smoothly? What does smooth mean? How can it do better?
The article that Tim sent me highlights that the harness is emerging as being as, and possibly way more, important than the parts/model(s) the agent is composed of. Therefore, choosing the right harness is crucial to a successful deployment of agentic AI. And comparing agents and their model stacks without including their harnesses is like comparing the aforementioned Formula 1 engines while ignoring what they are strapped onto.
When it's said out loud, it seems obvious that agentic AI must have governance. That does not mean we were ready to hear it before. As the article also points out, our focus has been on other aspects of AI up until now.
From 2022-24, we were all about the generative prompts and how to get them right. Last year was context-maxxing, that is, ensuring the models get the right information and continue to learn. And just now, in the first half of 2026, have we caught our AI bearings enough to start defining reasonable rules for engagement.
And it makes me happy because building, securing, and optimising the harness around the processing is a huge part of what we know how to do. So even when the agent technology is new and so much more powerful than anything deployed before, we have a proven methodology to get the most out of it.
Regitze Ladekarl, FRM, is FRG’s Director of Company Intelligence. She has 25-plus years of experience where finance meets technology.
This article is part of the FRG Risk Report, published weekly on the FRG blog. To read other entries of the Risk Report, visit frgrisk.com/category/risk-report/.