An AI agent pursues a goal, uses tools on its own and takes several steps in sequence. What separates it from a chatbot is not intelligence — it is the ability to act.
The term is currently applied to anything with a text box. It only becomes useful with a clean distinction.
A chatbot answers what you ask. An assistant completes a bounded task. An agent is given a goal and decides for itself which steps that requires — and when it is done.
Chatbot, assistant, agent #
- Chatbot — input in, answer out. One step, no memory beyond the thread, no tools.
- Assistant — completes a defined task, often with access to a system. "Put that meeting in my calendar."
- Agent — gets a goal rather than a task. "Make sure my calendar has two free mornings next week." It checks, plans, acts, verifies.
The boundary is fluid, and marketing blurs it further. The useful test: do I have to trigger every step myself? If yes, it is not an agent.
What an agent actually needs #
Four things, and all four are work:
- A clear goal — including a definition of when it has been reached.
- Tools — calendar, database, email. Without access, the best agent stays a text generator.
- Memory — what it has already tried and what failed.
- Limits — what it may never do without approval.
Unilateral actions: 0. My assistant may research, prepare and propose — but send nothing outward without my approval. That limit is not decoration; it is the condition under which I run something like this at all.
Where agents fail #
From my own setup with more than 180 installed agents: the problem is rarely the model. It is the environment.
Agents fail on vague goals, on tools that fail silently, and on missing stop conditions — an agent without one keeps trying forever. And they fail when someone applies them to tasks a simple automation would handle better.
Common questions #
What is the difference between an AI agent and a chatbot?
A chatbot responds to input, in one step and without tools. An agent receives a goal, plans several steps, uses tools such as calendars or databases, and verifies the result itself.
Does every company need AI agents?
No. Many tasks are solved more cheaply and reliably by classic automation. Agents pay off where the path to the goal varies by situation and decisions are needed along the way.
How many agents do you need?
Fewer than you think. Of more than 180 installed, I use maybe twenty daily, and five carry most of the load. The work is in orchestration, not collection.
Should agents act autonomously?
Technically they can; organisationally you should limit it. Anything with outward effect — sending, booking, publishing, paying — belongs behind human approval.
Sources #
- Anthropic — Building effective agents
- Regulation (EU) 2024/1689 (AI Act) — Art. 50(1): disclosure duty for AI systems interacting directly with people
- Own practice: setup and orchestration of 180+ agents, documented in "184 AI agents, one person"
Where does an agent pay off — and where not?
Ask me — or ask Ada. She is one herself, available around the clock, and will propose a slot with me if you need one. Free, no form.