Skip to content

AI agent

An AI agent is a system that uses information and tools to decide on and carry out steps towards a goal, with a defined degree of autonomy

What it means, how it is used and where it comes up, with its related concepts.

What it is and how it works

In LLM-based systems, the model can choose actions, observe results and adjust the next step. One example is reviewing a request, checking data and preparing a reply. Permissions, stopping conditions and actions that require human intervention must be defined. A chatbot that only answers questions does not necessarily work as an agent.

Also known as

Intelligent agent

Where we work on it

Technology

Frequently asked questions about AI agent

When is an AI agent better than an automated workflow with fixed steps?

An agent is worth considering when the path to completing the task changes depending on what it finds and it is not practical to define every step in advance. If the rules and routes are clear, a fixed workflow may be simpler to control. The choice should compare quality, time, required resources and ease of supervision for the specific task; more autonomy does not mean more usefulness.

Does an AI agent have to be allowed to take actions without review?

No. Autonomy can be limited to searching for information, preparing proposals or carrying out low-impact actions, while others are reserved for human approval. Permissions should be separated technically, and it should be defined when the system must stop or ask for help. In a hypothetical example, an agent could prepare an order change without confirming it. Supervision should match the risk and be workable in day-to-day work.

How do you evaluate an AI agent before adding it to a process?

You need to check both the task outcome and the path the agent takes to reach it. Include typical cases, incomplete information and tool failures, and observe whether it respects permissions and stopping conditions. A favorable demo is not enough. The evaluation should record errors, required interventions and resources consumed, so the agent can be compared with a simpler alternative to decide whether it adds value.

What happens if an agent finds instructions inside a document or a website?

It should treat that content as information to analyse, without automatically turning it into an authorised order. A document may contain instructions that try to divert the task or trigger data access and sending. It is worth separating the content consulted from the operating rules and controlling actions outside the model. Minimum permissions and review of sensitive operations reduce the reach of a failure, although they do not remove all risk.

How can you limit the damage of a wrong action?

By reducing what the agent can do and the number of operations it can run before a check. It is worth granting only the permissions needed, validating each request in the system that receives it and keeping useful logs to investigate errors. Where possible, actions should allow recovery. These controls must be applied technically: an instruction asking the model to behave carefully does not replace an access restriction.

When should an agent stop and escalate a task?

When it lacks essential information, cannot verify a result, comes across an action outside its permissions or reaches the agreed limits of attempts and resources. These conditions should be defined before putting it into use and checked during evaluation. The escalation should explain what has been done, what remains uncertain and what decision a person needs to make. Repeating steps without progress is not the same as solving the problem.

Related concepts

References