Two years ago, the best AI agents could complete an everyday task on a live website about 14% of the time. Today, the best systems are above 70%.
Better models are part of that progress, and so is the infrastructure around them. A lot of what makes an agent work happens outside the model, in what’s known as the harness.
I asked Pat Belliveau, our CEO at Gambit, to help me break down what that means.
What a Harness Is
“Think of the model as a brain,” Pat says. “But a brain without a body can’t do much. It can’t pick up a sweater. It doesn’t have arms or hands.”
The model does the thinking. The harness gives it those hands. It’s the structure around the model that connects it to tools and systems, manages the information it has access to, and puts rules around what it can do.
Giving the model the ability to act is one part of the equation. A good harness also gives that ability boundaries, and those boundaries depend on the kind of work the agent is doing.
“It’s like calling every car just ‘car,’” Pat says. “You can have a Corvette or a Honda. They’re both cars. That doesn’t tell you what either one can do.”
An agent that checks the weather every morning and sends you a message doesn’t need much around it. An AI worker operating inside business systems for hours at a time needs more structure around how it works. They’re both agentic, but the level of responsibility is different.
The Four Things a Harness Has to Handle
As Pat walked me through it, most of what a harness does fell into four areas.
What can the AI see?
The harness determines what information the AI has access to and where that information comes from. Access is defined around the work the agent needs to do.
Take Bogan, Gambit’s AI worker for buy-side deal origination. Bogan finds business owners who fit an investor’s thesis, researches and scores those opportunities, and helps get the first conversation started. The data it can work with is defined as part of that job.
What can the AI touch?
Seeing information and changing something are two different permissions. The harness defines which tools and systems the AI can use, along with what it’s allowed to do inside each one.
Fleet, Gambit’s freight-coordination worker, is a good example. It reads load documents, validates the details against the order, matches the carrier, and writes confirmed loads back to the transportation management system. Its access is built around the systems and actions required to move that process forward.
What can the AI decide?
Some decisions are routine enough for an AI worker to make on its own. Others need a person, and the harness draws that line before the worker starts running.
Chloe, Gambit’s voice concierge for the Town of Vail, answers resident questions using information from the municipality’s own systems. She can handle questions about things like parking, events, permits, and services, while requests outside her scope are routed to staff with the context from the conversation.
What happens when something breaks?
Even a well-defined process will eventually run into something unexpected, whether that’s a system becoming unavailable, information arriving in a format the worker can’t use, or a situation that falls outside its scope. The harness determines how the worker should respond when that happens.
Depending on the situation, the worker might retry, stop, flag the issue, or bring in a person. Those responses are designed into the system so there’s a defined path forward when the usual process doesn’t apply.
The model keeps the same underlying capabilities. The harness gives those capabilities a defined environment to operate within.
An AI Harness in Action
Pat gave me an example that makes this easier to picture.
Say a date was entered incorrectly across a batch of property records. An AI worker can go through those records and fix the date one at a time. The scope of the job is what makes this a useful example.
The worker stays focused on the field it was asked to update and the type of record it was given access to. It can keep moving through the task without asking for approval after every correction.
The harness has already established the scope of the job, including which records the worker can access, which field it can change, and what it should do if it encounters something unexpected. As long as the work stays inside those boundaries, it can keep going.
Those boundaries give the worker room to operate while keeping the job contained.
There’s still plenty of room for the tooling to improve. In Pat’s example, fixing records individually can be slower than making the same change in a batch. Better tooling can close gaps like that while preserving the boundaries around the work.
Reliability comes from defining those boundaries well and designing the system around the job.
Agent vs. Harness
“Agent” and “harness” sometimes get used as though they mean the same thing. In reality, they describe different parts of the system.
An agent is software that can take actions without being prompted for every individual step. The harness is the structure around it that determines how those actions happen.
For example, something that checks the weather every morning at 8 a.m. and sends you a message is agentic. It can run on a schedule and complete its task without someone prompting it in that moment. The job is narrow, so the surrounding structure can stay relatively simple.
The requirements grow when an agent needs to update a customer record, book something, change a status, or take some other action that sticks. As the agent takes on more responsibility, defining what it can access, change, and decide becomes more important. It needs a harness.
That’s also why we use the term “AI worker” for what we build. A worker is built around a specific job, and the harness follows from that job. It defines what the worker needs to see, what it needs to touch, what it can decide, and where a person still needs to be involved.
A Few Quick Answers
What is an AI harness?
An AI harness is the structure built around a model that lets it take action inside a defined environment. It manages what the AI can see, which tools it can use, what it can decide, and what happens when something goes wrong.
What’s the difference between an AI agent and an AI harness?
An agent is an AI system that can take actions without being prompted for every individual step. A harness is the structure around the agent that controls how and where those actions happen.
Why do AI agents need a harness?
A harness gives the AI defined boundaries, permissions, and rules for how to operate inside the systems it uses. That structure makes it possible to give the agent more responsibility while keeping the work within a defined scope.
What makes a good AI harness?
Clear information boundaries, defined tool access, decision limits, and reliable failure handling. The right version of each depends on the job the agent is being asked to do.
Giving AI Responsibility
A good harness gives an AI worker the structure it needs to be useful inside a business. That means giving it the right information and tools, deciding what it can handle on its own, setting boundaries around the work, and defining what should happen when it reaches one of them.
When we build an AI worker at Gambit, we start by asking what would have to be true for us to trust it with the job.
If you’re trying to answer that question for your business, let’s talk.
