If you feel like artificial intelligence has changed five different times in the last few years, you're not wrong. It kind of has.
Let's rewind to November 2022. OpenAI released ChatGPT to the public, and suddenly millions of people were having full conversations with AI for the first time. You could ask it to write an email, explain a complicated topic, come up with a business idea, write code, or make up a bedtime story about a dinosaur going to space.
The possibilities were endless. And it felt pretty dang incredible.
For a while, there was this idea that software as we knew it might be on its way out. If AI could understand what you wanted and just do it for you, why would we need all of these different programs, buttons, forms, and workflows?
Then people started trying to use AI for work, and things got a little more complicated.
AI was impressive, but it was also unreliable. It could write a great email and then confidently make up a fact two sentences later. It could explain how to update your CRM, but it couldn't necessarily go into your CRM and do it. And even when it eventually gained access to tools, getting it to reliably complete a long process from beginning to end turned out to be much harder than giving it a good prompt.
A lot has changed since then. The models have gotten much better, but something else has happened too. We've gotten much better at building the systems around them.
2023: Why Early AI Couldn't Access Your Business
A good way to think about those early AI systems is to imagine a really smart person sitting in a room with no windows.
Before they went into the room, they read an enormous amount of information. Books, websites, articles, code, conversations, and all kinds of other material. Because of that, they can answer an incredible number of questions.
Ask them to explain how mortgages work and they'll probably do a pretty good job. Ask them to write a funny poem about your dog and they'll do that too. But ask, "Did this customer pay their invoice this morning?" and suddenly they're stuck.
That information isn't something they learned before going into the room. It's sitting inside your accounting software right now, and they can't see it.
The same thing happens if you ask about your company's internal refund policy, this morning's sales numbers, or what happened in a customer conversation yesterday. Unless you give the AI that information somehow, it doesn't know.
This was also part of the reason hallucinations became such a big issue. Large language models are designed to generate likely responses, not look things up inside some perfect database of facts. Sometimes that means they produce something that sounds completely reasonable but happens to be wrong.
People tried to solve a lot of this with better prompts. "Don't make anything up." "Only give me information you're sure about." "Double-check your answer before responding." Those instructions can help, but they can't give AI access to information it doesn't have.
No matter how nicely you ask the person in the room, they still can't tell you whether an invoice was paid if they can't see the accounting system.
2024: How RAG Gave AI Access to Your Information
One of the big developments that helped with this was something called Retrieval-Augmented Generation (RAG). It sounds much more technical than it is.
Think about our person sitting in the room again. This time, imagine we put a filing cabinet beside them filled with your company's documents. When you ask a question, a system searches through those files, finds information that might be relevant, and hands it to the AI before it answers you back.
Now if you ask, "What's our refund policy?", the AI can read your actual refund policy instead of trying to answer based only on what it already knows. That's basically the idea behind RAG.
It made AI much more useful for businesses because suddenly the model could work with internal documents, policies, training materials, product information, customer data, and other information that wasn't part of its original training.
It didn't make AI perfectly accurate, though. That's an important distinction.
Research on RAG systems has found that they can still hallucinate because there are multiple places where things can go wrong. The system might retrieve the wrong information, miss something important, or the model might misunderstand the information it was given. RAG helps ground an answer in real information, but it doesn't guarantee the answer will always be right. A 2025 review published by ACM goes into these limitations in much more detail.
Even with that limitation, this was a huge improvement. AI could finally know much more about your actual business. But businesses quickly ran into another problem.
Knowing what to do isn't the same as doing it.
Imagine an AI that can read your entire customer support manual. A customer asks for a refund, the AI checks the policy, looks at the situation, and tells your employee, "Yes, this customer qualifies." That's helpful, but someone still has to open the billing system, find the transaction, issue the refund, update the customer record, and send the confirmation.
At some point, businesses stopped asking, "How do we get AI to answer better questions?" and started asking, "Can it just handle the work?"
2025: How AI Agents Started Taking Action
This is where AI agents really started getting attention.
Instead of giving AI access only to information, developers started giving it access to tools too. That could mean browsing a website, searching a database, sending an email, running code, updating software, creating a file, or calling another system.
Suddenly AI wasn't limited to telling you what to do. It could start doing things itself.
For example, instead of asking AI to tell you how to research three competitors, you could potentially give an agent the goal of researching those competitors, gathering the information, comparing them, and putting the findings into a presentation.
That was a massive shift, and some of the demos looked unbelievable. The problem was that real work tends to be messier than demos.
Think about a customer support request. An AI agent might need to read the message, figure out who the customer is, find their account, understand what they're asking for, check company policy, decide what action to take, update a system, and send a response. That's already a lot of steps.
If the AI makes a mistake near the beginning, that mistake can follow it through the rest of the process. It could identify the wrong account and then correctly perform every remaining step on the wrong customer. Or it could misunderstand a policy and then very efficiently take the wrong action.
This became especially noticeable with longer tasks. Early agent systems could be very impressive when everything went according to plan, but unexpected situations were much harder. A page might change, a tool might fail, information could be missing, or, the system might need to decide whether to try again, take a different path, or stop completely.
One of the early benchmarks that showed how difficult this was is called WebArena. It tests AI agents on realistic website tasks, like navigating pages and making changes across different systems. In the original research, the best GPT-4-based agent completed only 14.41 percent of the tasks successfully, compared with 78.24 percent for humans.
That number has improved dramatically since then. The Stanford AI Index 2026 reports that top performance on WebArena reached 74.3 percent in early 2026. That gives you a sense of just how quickly this technology is improving.
But better models aren't the whole reason these systems have gotten more useful. We've also learned a lot about how much structure AI needs around it.
2026: What Makes AI Agents More Reliable?
For the last few years, most AI conversations have focused on the model itself. Which one is smartest? Which one has the biggest context window? Which one scored highest on a new benchmark?
Those things matter. But if you're trying to give AI responsibility inside a business, the model is only one part of the system.
A really smart model with access to your business still needs rules. It needs to know which tools it's allowed to use, which information it can access, what a correct result looks like, what to do when something fails, and when it needs to stop and ask a person for help.
This is where the idea of an AI harness becomes useful.
A harness is basically the structure around the AI that helps it work reliably. You can think about it a little like hiring a very capable new employee. You wouldn't give someone a laptop, all of your company passwords, and say, "You're smart. Go and figure it out."
You'd give them a job description. You'd train them on your processes. You'd give them access to the systems they need. You'd tell them what they're allowed to approve themselves and what needs a manager. You'd probably check their work more closely at first. And you'd make sure there was a process for dealing with situations they hadn't seen before.
AI needs a lot of the same things.
OpenAI wrote about this in a 2026 article on harness engineering. Their team used AI coding agents to build an internal software product containing roughly one million lines of agent-generated code. The project also showed how much infrastructure was needed around the agents to make that level of output reliable.
The team built tests, documentation, tools, feedback loops, validation systems, and rules that helped the AI understand what good work looked like. OpenAI describes the engineers' role as shifting away from writing code toward designing the environment and feedback loops that allowed agents to work reliably.
OpenAI has since used the term "model-native harness" for infrastructure that gives agents controlled access to files and tools, keeps track of state, and lets them continue from checkpoints when something goes wrong.
A lot of the progress we're seeing now comes from improving that surrounding infrastructure. Better models obviously help, but businesses also need systems that give those models the right context, boundaries, tools, and checks before they can be trusted with work.
How Does an AI Harness Work?
Let's say you want an AI worker to handle freight load confirmations.
You could technically connect an AI model to your systems and tell it, "Handle incoming load confirmations." But there's a huge amount of information hiding inside that sentence.
- Which loads is it allowed to confirm?
- What information has to be present first?
- Which rates are acceptable?
- What happens if paperwork is missing?
- What should it do if something is outside the normal policy?
- Which systems can it update?
- What happens if one of those systems goes down?
A dependable AI worker needs answers to all of those questions.
So instead of giving the AI unlimited freedom, you build boundaries around the job. The worker gets access to the information it needs, like the incoming request, load information, rates, and company policies. It gets specific tools for checking records, updating the system, and sending confirmations. It also gets rules around when it can act on its own and when it needs to involve someone from the team.
You can also check parts of its work automatically. If ten pieces of information are required before a load can be confirmed, the system can make sure all ten are there. If a number needs to fall within a certain range, software can check that before anything gets approved. If something doesn't make sense, the workflow can stop instead of letting one bad decision turn into five more.
The same applies to longer jobs. Important information about what has already happened can be saved outside the AI itself, so if something fails halfway through, the system doesn't necessarily need to start all over again.
And maybe most importantly, the AI needs to know when something isn't its decision to make.
If an AI usually handles routine refunds under $100 and suddenly receives a request for $25,000, it shouldn't matter that the AI knows how to press the refund button. That's outside its job. The system should stop and ask the right person.
All of these pieces work together to give the AI structure around the job it's responsible for. It knows what information it can access, which actions it can take, how its work gets checked, and when something needs to be handed back to a person. That's really what we mean when we talk about a harness.
What's the Difference Between an AI Agent and an AI Worker?
This is also where we think there's an important difference between an AI agent and an AI worker.
An agent can take action. A worker has a job.
That might sound like a small distinction, but it changes how you build the system.
At Gambit, every worker starts with a specific outcome. A logistics company doesn't need "an AI agent." It might need loads confirmed. An investment firm might need qualified opportunities found. A municipality might need resident questions answered. A hotel sales team might need new prospects identified and contacted.
The technology underneath those workers can include AI models, agents, retrieval systems, workflows, traditional software, and other tools. But those aren't really the thing the business cares about.
The business cares about whether the job gets done.
That's why we've become much more interested in building AI around clearly defined jobs instead of simply giving businesses another AI tool to use.
A chatbot still needs someone to ask it questions. A copilot helps someone do their job. An agent can take action on someone's behalf. An AI worker is designed to own a specific piece of work from beginning to end.
That's a very different relationship with the technology.
What Does This Mean for Businesses Using AI?
If you've experimented with AI over the last few years and thought it was impressive but not quite reliable enough to trust with real work, that's a pretty reasonable conclusion.
The technology you were using may have genuinely not been ready for the job you were asking it to do.
What's changing now is that businesses don't necessarily have to choose between AI that only gives advice and AI that gets unlimited freedom to do whatever it wants. There's a much more practical middle ground where AI can own clearly defined work inside boundaries that are designed around the business.
The models are better. The tools are better. But the systems around them are getting better too, and that's arguably what makes the biggest difference when you move from playing with AI to actually putting it to work.
That's the idea behind the AI workers we're building at Gambit. Each one has a specific job, access to the information and tools it needs, clear boundaries around what it can do, and a defined outcome it's responsible for.
For businesses, that opens up a much more practical way to think about AI. Instead of asking where another AI tool might fit, you can start looking at the work already happening across your business and asking which parts of it AI could realistically own.
If you're starting to think about what that could look like inside your business, let's chat.
