What is an AI agent?

Erindreki is the Icelandic word for what English calls an agent. In short, an agent is AI that is given a goal and finds its own way to reach it — instead of being told exactly what to do at every step.

An agent is not the same as a chatbot

A chatbot answers questions in a conversation. It waits for you to ask and answers what you asked. An agent goes further: it takes on a task, decides which steps are needed, uses tools to carry them out, and keeps going until the goal is reached — or until it needs to stop and ask a person for help.

That’s why we keep the two apart on erindrekar.is: agents and chatbots each get their own listing, even though both are built on the same underlying technology.

Three parts: thinking, tools, observation

Hugging Face’s Agents Course describes the core of an agent as a repeating loop: the agent thinks about the next step, acts using a tool, and then observes the result before deciding on the next step. This loop — thought, action, observation — repeats until the task is done.

Lilian Weng, who has written extensively about autonomous agents built on large language models, describes a similar structure built on three pillars:

• Planning — the agent breaks large tasks into smaller steps and can reassess its own decisions afterwards.
• Memory — the agent uses both the context it has right now (comparable to short-term memory) and, in many cases, can retrieve older information from a database (comparable to long-term memory).
• Tool use — the agent uses external tools and systems (e.g. APIs, databases, search engines) to do what a language model alone could not.

When to use an agent — and when not to

In its article on building effective agents, Anthropic draws an important distinction: on one side are workflows, where the steps are fixed in code in advance, and on the other are agents, where the AI itself directs how it solves the task and which tools it uses along the way.

The advice is simple: start with the simplest solution that works. Workflows suit predictable tasks with clear steps. Agents are the right choice when the task is open-ended and it’s impossible to say in advance how many steps are needed, or which path is best. The more complex and autonomous the agent, the more it matters to design its tools well and test it thoroughly — because mistakes can compound when an agent makes many decisions in a row without oversight.

Agents are a business matter, not just a technology project

In an opinion piece in Viðskiptablaðið, Hlöðver Þór Árnason, CEO of APRÓ, points out that companies use AI in practice at three levels:

1. Individual use — employees use AI for specific tasks, often with single agents that each person sets up for themselves.
2. Teams share the gains — teams, for example software teams, build shared ways of working with AI, but the knowledge often stays stuck in silos.
3. The company builds a holistic system — agents are coordinated across the company and run with a process-driven approach, led by the business side, not by the IT department alone.

The core of this view is that agents deliver the most value when they are part of a clear company strategy — not just a technical experiment. That’s why the goal of erindrekar.is is to gather Icelandic material and real examples of agents and MCP servers that Icelandic companies can learn from and use.