Short answer: an AI agent is a language model that does not just answer, but takes steps on its own to finish a task. It uses tools, such as searching, opening a file or filling something in a system, and checks the result after each step.
Chatbot, workflow and agent
In its widely read piece Building effective agents, Anthropic draws a useful line:
- Workflow: the language model and tools follow a fixed route that a developer laid out in advance. For example: summarise an incoming email, decide whether it is a complaint or a quote request, and put it in the right queue.
- Agent: the language model decides for itself which steps are needed and which tools to use. It keeps going until the task is done or it gets stuck, and can come back to a person.
An ordinary chatbot does neither. It answers and waits for your next question.
When does an agent make sense?
Anthropic recommends starting with the simplest solution and only adding complexity when it clearly works better. Often one good request to a model, with the right information attached, is enough. A fixed workflow is more predictable. An agent fits tasks where you cannot know in advance how many steps are needed.
Two areas Anthropic names where agents work well:
- Customer support, where the agent looks up customer data and order history and takes actions such as updating a ticket.
- Coding, because you can test the result.
What to watch out for
- Costs and errors add up. An agent takes many steps, and each one costs computing power. An early mistake can carry through. Anthropic advises testing agents extensively in a sandboxed environment and building in limits, such as a maximum number of steps.
- Keep a person in the loop for actions that matter, such as sending, paying or deleting.
- Give an agent access only to what it needs. The Dutch Data Protection Authority points out that an AI system with access to your files should not see more than the employee using it.
Sources
- Anthropic: Building effective agents (December 2024)
- Autoriteit Persoonsgegevens: Hulpmiddel inzet generatieve AI & de AVG (PDF, Dutch, on access control)
Accessed 21 October 2026.