A definition that is actually useful
An AI agent receives a goal, not a command. It decides for itself which steps to take, calls external tools, looks at the result and adjusts its behaviour until it reaches the goal or runs into the limit of its authority.
A handy formula for checking: goal → plan → action → observation → memory → adjustment. If a system lacks at least observation and adjustment, what you have is a text generator in a wrapper.
How an agent differs from neighbouring concepts
From a chatbot
A chatbot answers. An agent acts: it files a ticket, finds a document, recalculates, sends the result to your system. An agent may have no chat window at all: it can work on an event, without messages.
From classic automation
An automation script runs a sequence described in advance. An agent builds the sequence itself for the case at hand. On varied inputs that is its strength. Where exactly the same order of actions is required, it is its weakness.
From “a model with a prompt”
A prompt sets behaviour within one exchange. An agent lives longer: it remembers the context of the task, sees the results of its actions and is responsible for the final result as a whole.
What it is made of
- The model decides on the next step. The same agent architecture can run on different models.
- Tools give access to databases, APIs, search and code. Through them, reasoning becomes action.
- Memory holds the context of the task, the history of steps and long-term knowledge about the domain and the user.
- The planner breaks the goal into steps. In simple agents the model itself plays this role.
- The control loop sets the limits of authority, the points where a person confirms, and the action log.
- Evaluation: a set of test tasks and metrics. Without them you cannot tell whether things got better.
Which tasks to give it
A good candidate for an agent meets four conditions at once:
- The task repeats. Otherwise preparing the tools and checks will not pay off.
- The result can be checked without relying on an impression.
- The inputs are varied. Otherwise an ordinary automation script is cheaper.
- The cost of an error is known. It determines how much control to build in.
Bad candidates: one-off tasks, processes with no criterion of correctness, and operations with irreversible consequences where a person cannot confirm the action.
Autonomy is switched on in steps
A sensible rollout goes in steps, and the next one is switched on only after the metrics have confirmed the previous one:
- The agent proposes, a person executes. This already brings value: you see the quality without risk.
- The agent executes, a person confirms every action.
- The agent executes on its own, and a person confirms only risky operations.
- The agent works autonomously, and a person handles exceptions and reads the reports.
Jumping straight to the fourth step is what most often forces pilots to be rolled back: without accumulated statistics, nobody can say what exactly the system is doing wrong.
What to measure
- the share of tasks carried through to the end without a person stepping in;
- the share of results that had to be corrected;
- the cost of one solved task, not of one model call;
- the time from assignment to result, compared with the manual process;
- the share of correct refusals, when the agent rightly decided the task was not its job.
A common implementation mistake
Starting by choosing a framework. It is more useful to describe one scenario first, down to the level of “what counts as a successful result”, and to check that the agent has access to the data and systems it needs at all. The framework can be changed later, while a wrongly chosen scenario eats up the pilot budget.