Last spring, a support team at a mid-sized software company logged on to find their overnight queue already sorted. Tickets that arrived at 2 a.m. had been tagged, categorised, and in about a third of cases answered. Nobody had been on shift. The work was done by software that read each ticket, worked out what kind of problem it was, pulled the customer’s billing history from an internal API, drafted a reply in the company’s tone, and pushed anything unusual into a human’s inbox for morning review.
That software is an AI agent. Not a chatbot waiting for you to type, and not a language model that answers one question and stops. An agent takes a goal, splits it into steps, uses tools, checks its own output, and keeps going until the job is finished or it decides it needs a person.
The Definition, Without the Marketing
An AI agent is a system that uses a large language model as its reasoning engine and wraps it in three things the model does not have on its own: the ability to act, a memory of what it has already done, and a loop that lets it keep working across many steps.
That last part is the one people miss. A single model call is a snapshot. You send text, you get text back, and the model forgets you exist. An agent run is a process. It might make forty model calls before it is done, and each one builds on the results of the last.
The four pieces you will find in almost every agent
- The model, which handles reasoning, planning, and deciding what to do next.
- Tools, which are how the agent touches the outside world: a search API, a database query, a code interpreter, a payment system, a browser.
- Memory, both the short-term scratchpad for the current task and any longer-term store of facts, preferences, or past outcomes.
- A control loop, the logic that decides when to keep going, when to try a different approach, and when to stop and ask for help.
Strip out any of those four and you are back to something that looks more like a chatbot with extra steps.
The loop, in plain English
Almost every agent, from a weekend hobby project to a production system chewing through mortgage applications, runs some version of the same cycle. It observes the current state. It decides what the next useful action is. It takes that action. It looks at what happened. Then it goes round again.
Simple to describe, hard to make reliable. Small mistakes compound. If step three of a twenty-step task goes slightly wrong, the agent can spend the next ten steps confidently building on a bad assumption.
Agent vs. Chatbot vs. Copilot
These three get lumped together constantly, and the difference matters when you are deciding what to buy or build.
A chatbot responds. You ask, it answers, the interaction ends. A copilot suggests while you work: autocomplete in an editor, a draft email you still have to send. An agent acts. It holds the goal rather than the next sentence, and it can finish a task end to end without you approving every move.
That shift from suggesting to acting is what makes agents useful and what makes them dangerous. Autocomplete that gets a word wrong costs you a keystroke. An agent that applies the wrong refund policy forty times before anyone notices costs you customers.
What AI Agents Are Actually Doing Right Now
The clearest traction is in software development. Coding agents read a repository, write a patch, run the test suite, and iterate on failures. Plenty of teams now treat them as a junior contributor who works overnight. But more code is not the same as better software, as this look at why AI coding agents generate more code but not more software makes clear. Review capacity, not typing speed, is usually the bottleneck.
Outside engineering, agents have crept into back-office work that used to be pure human grind. Mortgage lending is a good example. A loan file involves income verification, title checks, and a pile of documents that all have to agree with each other. Vesta raised $30M to bring swarms of agents to mortgage lenders, betting that coordinated groups of them can work through that paperwork faster than a human team, with a person signing off on anything that smells off.
Sales research, customer onboarding, invoice reconciliation, and internal IT requests are getting the same treatment. The buyers are no longer just developers, either. Nous Research hitting a $1.5B valuation and launching agents aimed at business users tells you where the money thinks this is heading.
Why You Keep Hearing About Multi-Agent Systems
One agent trying to do everything gets muddled. A common fix is to split the job across several specialised agents that hand work to each other: a planner, a researcher, a writer, a reviewer. Each one carries narrower instructions and a smaller pile of context.
This pattern is old enough to have tutorials. You can build a small Python tool with a team of AI agents using MetaGPT in an afternoon and watch a product manager agent, an engineer agent, and a QA agent argue their way to working code. It is a genuinely useful exercise, mostly because it shows you how often the handoffs break.
Swarms sound impressive. In practice the coordination overhead is real, and adding a fifth agent to a four-agent system often makes things worse before it makes them better.
Where Agents Still Fall Over
Error compounding
This deserves its own heading because it is the number one reason agent pilots die. Long-horizon tasks need checkpoints, verification steps, and a way to roll back. An agent that can undo its last action is worth far more than one that is simply smarter.
Security and control
An agent with access to your inbox, your files, and a payment API is a powerful thing to hand to a system that can be talked into things. Safety teams are treating it seriously. Anthropic’s response was blunt: after failing to reliably control its own agents, it cut its internal evaluation systems off from the live internet. If you are deploying agents with real permissions, budget for oversight the way you would budget for the agent itself.
Evaluation is harder than it looks
With a chatbot you can eyeball fifty responses and form a view. With an agent that takes two hundred steps to finish a task, did it do well is a genuinely hard question. You need traces, step-level scoring, and someone willing to read the logs.
Is Your Task a Good Fit for an Agent?
Run the job through this filter before you reach for one.
- Can you define done? If you cannot write a pass or fail check, the agent cannot either.
- Does it need tools? Agents shine when the work involves lookups, calculations, or changes to another system. Pure text generation rarely needs one.
- Is the volume high enough? Building and babysitting an agent costs real effort. Ten tasks a month probably is not worth it.
- What happens when it is wrong? Reversible mistakes are fine. Irreversible ones need a human gate.
- Can you see what it did? Without logs and traces you cannot improve the thing, and you cannot defend it when a customer asks.
The Gap Between a Demo and Something That Ships
Watching an agent book a flight in a five-minute video is fun. Getting one to process four hundred insurance claims a day without a single one going sideways is a different sport, and it is where nearly all the real work lives.
The teams getting value out of agents today tend to share a few habits. They scope the task tightly, so the agent is not reasoning about the entire business. They give it good tools and clean data, because an agent reading a stale CRM is just a faster way to be wrong. They put a person at the decision points that actually matter and automate the twenty boring steps on either side. And they measure the agent’s output the same way they would measure a new hire’s, with sampling, spot checks, and a willingness to switch it off.
The technology is moving quickly enough that the specific tools in this article will look dated within a year. The questions worth asking will not. What is the goal, what can it touch, how do you know it worked, and who catches it when it does not.