Seriously, what is an Agent?
A lot of nonsense gets sold. We are changing that. Part of the work is clean definitions.
To a user, an agent feels like ChatGPT: I type something in and get an answer back. But ChatGPT and the like feel like a helpless brain: you constantly copy information back and forth and learn a great deal, while the actual execution still falls to the human.
In principle, an agent is exactly that kind of brain - with hands and a plan.
Based on the user’s input, the AI model in the background builds a plan. Then the decisive part: the plan is executed, it almost always involves access to various applications, and during execution the brain automatically checks whether everything is going according to plan.
The definition becomes clearest when we look at different prompts. Anyone who has already worked with “Deep Research” in ChatGPT or Gemini has already encountered a powerful but internally very simple agent (its only tool: web search).
What agents can do
A chatbot answers questions. An agent is given tasks. That makes sentences possible that you could not say to a computer before:
The difference: autonomy
The technically decisive thing about an agent is its autonomy. With classic software, a developer specified every step. An agent decides for itself which steps a task needs and in which order. Anthropic draws exactly this line: a fixed workflow follows predefined paths, an agent determines its own.
On top of that: because an agent works in many small steps, it can think a problem through longer and more thoroughly than a single chat reply. It works in a loop.
Automation ≠ agent
Zapier, Make.com or n8n are not agents but automations, where you connect boxes with lines and fill in configurations. An agent, by contrast, you simply talk to.
- Anthropic — Building Effective Agents (2024)
- Google Cloud — What are AI agents?
- Model Context Protocol — open standard for tool use by agents