The word "agent" is everywhere in AI right now, often without definition. At its core an agent is a language model given a goal, a set of tools, and permission to take steps on its own until the goal is met.
01From answering to acting
A chatbot is reactive: you ask, it answers, the exchange ends. An agent is goal-directed. Tell it "find three suppliers and draft an email" and it will break that into steps, search, read, decide and write — looping until done, without a prompt for each step.
02What makes an agent
Four ingredients turn a model into an agent. A goal to pursue, the ability to plan a sequence of steps, tools it can call (search, code, a calculator, an API), and memory to carry context across steps. Remove any one and you have a chatbot again.
A language model supplies the reasoning; these four capabilities turn that reasoning into action over time.
03The agent loop
Agents run a cycle often summarised as "think, act, observe". The model reasons about what to do next, calls a tool, reads the result, and decides whether the goal is met or another step is needed. It is the same closed loop as reinforcement learning, but driven by language reasoning rather than numeric reward.
The loop repeats until the goal is reached or a limit is hit. Each pass grounds the next decision in a real result rather than a guess.
04Tools & memory
Tools are what let an agent affect the world beyond text — searching the web, querying a database, running code, sending a message. Memory lets it remember what it already tried so it does not loop forever. The art is giving an agent enough tools to be capable but not so much reach that a mistake becomes costly.
05Where agents help
Agents shine on multi-step tasks with a clear goal and checkable results: research and summarisation, data gathering, routine workflow automation, customer triage, and coding assistance that runs and tests its own output. They struggle where goals are vague, errors are irreversible, or success cannot be verified.
06The real risks
Autonomy cuts both ways. An agent can take a wrong action confidently, be manipulated by malicious content it reads (prompt injection), rack up cost in a runaway loop, or exceed its intended scope. The mitigations are practical: least-privilege tool access, human approval for irreversible actions, spending and step limits, and full logging. Capability without guardrails is the mistake to avoid.
What to remember
- An agent is a language model with a goal, plus planning, tools and memory.
- Chatbots react; agents pursue a goal across multiple steps.
- The engine is a loop: plan, act, observe, repeat until done.
- Tools let agents affect the world; memory keeps them on track.
- They suit multi-step tasks with verifiable results.
- Guardrails — least privilege, approvals, limits, logging — are essential.
We translate advanced technology and EU policy into practical training. This explainer is part of our open Insights series for educators, youth workers and SMEs.