From answering to doing
Ask a chatbot to book a meeting room and it will tell you how to book a meeting room. Ask an agent and it will check the calendar, find a free slot, make the booking and send the invite. That gap, between producing text about a task and actually carrying it out, is what people mean when they talk about agentic AI.
The language model underneath can be exactly the same in both cases. What changes is the software built around it: a loop that lets the model decide on a step, take it, look at the result and decide again.
What an AI agent actually is
An AI agent is a program that is given a goal rather than a single question. To reach that goal it breaks the work into steps, calls whatever tools it has been given (a search API, a database, a code interpreter, an email client) and keeps going until the goal is met or it runs out of options.
Three things set it apart from an ordinary call to a language model:
- It plans. A request like "summarise last quarter's support tickets and flag anything urgent" turns into several smaller tasks: fetch the tickets, group them, read the urgent ones, write the summary.
- It acts. The model can call tools, and the results of those calls change what it does next.
- It corrects itself. If a query fails or returns nothing useful, the agent tries a different approach instead of handing the error back to you.
How it works under the hood
Whatever framework they are built with, most agents today follow the same basic cycle.
- Plan. The model reads the goal and decides what to do first. Some systems write out a full plan up front; others decide one step at a time.
- Act. The model picks a tool and fills in its arguments. The surrounding code, not the model, executes the call. That detail matters a great deal for security.
- Observe. The tool's output goes back into the model's context. Was the result what it expected? Is the task done?
- Remember. On longer tasks the agent keeps notes: what it has tried, what worked, what the user prefers. That memory can be a simple running log or a vector database.
The loop repeats until the model decides it is finished. Because every step depends on the one before, an agent's behaviour is settled at runtime. You can't read the code and know exactly what it will do, and most of the engineering effort goes into living with that fact.
Where agents are already useful
The best uses so far have one thing in common: the task is made of many small steps, and each step is easy to check.
- Software development. Coding agents read a repository, make a change, run the tests and fix whatever breaks. The test suite gives them a clear signal of success.
- Operations. An agent can watch metrics, notice an anomaly, pull the relevant logs and draft a diagnosis for the engineer on call.
- Research and analysis. Querying a data warehouse, comparing numbers and writing up what changed is tedious for a person and a good fit for an agent, as long as someone checks the conclusions.
Autonomy needs limits
An agent that can send email can send the wrong email. Once a model is allowed to act, a mistake is no longer just a bad answer. It is something that has already happened. Well-built agent systems therefore share a few habits:
- Narrow permissions. The agent gets the smallest set of tools it needs, read-only wherever possible, and risky code runs in a sandbox.
- A record of every step. Each tool call is logged with its arguments, so a person can see afterwards why the agent did what it did.
- A human checkpoint. Deleting data, spending money or contacting customers waits for someone to approve it.
What comes next
For many agentic tasks the models are already good enough. The hard part now is the engineering around them: keeping long tasks on track, testing systems that behave differently from one run to the next, and deciding how much freedom to give them. Teams that treat an agent as a distributed system with one unreliable component, rather than as a clever assistant, tend to end up with systems they can trust.
Next in the series: Part 2 compares agentic AI with generative AI and looks at what the difference means for chatbots.
