What is an agent, actually?
An agent is a software harness driven by a large language model. That's the core of it. Everything else — chatbots, coding assistants, workflow automation tools — sits on top of that basic structure. The label changes depending on what the harness is built to do, but underneath they're all the same thing: a language model with a layer of software around it that lets it act in the world.
How does it work?
Think of a road vehicle. The language model is the engine. Everything else — the chassis, the wheels, the controls, the body — is the agent. You can build that vehicle for different terrains and different purposes. A sports car and a lorry have very different builds, but they both need an engine to move. Agents work the same way. The language model provides the reasoning; the agent determines what it can do with that reasoning, what tools it has access to, what guardrails it operates within, and what environment it's built for.
This matters because when people ask "which language model should I use?" they're really asking which engine to put in the vehicle. It's an important question, but it's not the only one.
What can you expect it to do?
That depends almost entirely on how the agent is built. Set up correctly, an agent can do almost anything in the digital world — search, read, write, send, retrieve, analyse, decide, and act across systems and platforms. The better question isn't what an agent can do in theory. It's what do you need your agent to do? Start there, and build the agent around that answer. A generic agent built for everything is usually excellent at nothing.
What does a bad agent look like?
A bad agent is one that isn't fit for its environment. That can mean it's been given the wrong tools, the wrong instructions, or too broad a mandate. But it can also mean something more serious. Agents that aren't properly secured can be a vulnerability — they can be manipulated, exploited, or used in ways their owners never intended. An agent operating in an environment it wasn't designed for is unpredictable. And an unpredictable agent in a business context isn't a minor inconvenience. It's a liability.
What should a good agent do differently?
Three things. First, it should be competent in the environment it's placed in — the right tools, the right scope, built for the task it's actually being given. Second, it should stay in its lane. An agent that does things it wasn't asked to do isn't being helpful. It's being unpredictable. Third, everything it does should be traceable. You should be able to look at what an agent did, when it did it, and why — not just in theory, but in practice, with a clear record you can actually read.
Competence, restraint, traceability. If your agent doesn't have all three, you don't fully control it.
Should your business have one?
If your business has online tasks and functions that are repeated, time-consuming, or error-prone, an agent could help you automate them reliably. That's the honest answer.
But here's where most businesses go wrong: they look at the agent first. The right place to start is your own business. What do you actually need help with? Where are the bottlenecks? Where does productivity break down? Answer those questions first, then look at whether an agent is the right tool to address them.
An agent built around a clear business need is a force multiplier. An agent acquired because the technology is exciting is an expensive experiment.
Agents are not magic.
They are software — powerful, flexible software — that reflects the quality of the thinking that went into building them. The businesses that will get the most from them are the ones that approach them like any other serious operational decision: clearly, deliberately, and with an honest view of what they actually need.