What exactly is an AI agent?

"AI agent" has become the default label for almost any AI-powered product released in 2026. The term covers everything from a chatbot with a few API calls to a fully autonomous system that operates across an enterprise stack. The stretch makes it hard for decision-makers to compare offers.

This article sets out what an AI agent actually is — and what it isn't. It uses Anthropic's working definition, distinguishes agents from their neighbours (chatbots, workflows, RPA), names the four building blocks, and explains where agents fit and where they fail. The closing section covers what this means for German companies under DSGVO.

Definition

An AI agent is software you give a goal to. It works out the steps, calls the tools, looks at what came back, and adjusts. It keeps going until the goal is met or until a guardrail you set stops it.

Anthropic draws a clean line between two patterns. A workflow is a system where you wrote the steps and the model fills them in. An agent is a system where the model picks the steps itself. Sam Altman has called "agent" a weakly defined term, which is fair. The dividing question is who chooses the next move: your code, or the model.

Distinguishing agents from chatbots, workflows, and RPA

A chatbot replies to messages. Useful, but it does not go and do things in your other systems.

A workflow runs a fixed sequence. "Inbound mail → classify → if invoice, extract → push to ERP." The LLM might do the classifying. The path is yours. Nothing branches outside what you wrote.

RPA imitates a person clicking through a UI. Quick, dumb, brittle the moment the interface shifts a pixel.

An agent holds the goal, not the path. Give it "find the replacement part for this customer's broken pump in our three supplier catalogues" and it queries each catalogue, reconciles different SKU formats, and returns an answer with its reasoning. The steps for a hydraulic seal are not the steps for a control board, and you didn't have to write either set.

The four building blocks

A model does the thinking. Claude, GPT, Gemini, or a locally hosted Llama, Mistral, or Nemotron. Marketing slides argue about model choice. In production it usually matters less than the rest of the stack.

Tools are the hands: APIs, search, code execution, file access, database queries. In 2026 these increasingly go through the Model Context Protocol (MCP), the open standard Anthropic published in late 2024. OpenAI and Google DeepMind now back it, and governance moved to the Linux Foundation in December 2025. MCP gives every agent the same way to reach into your CRM. Until it existed, every project reinvented that wiring.

Context is what the agent has loaded for the current task: system prompt, conversation, retrieved documents, prior steps.

Guardrails are the part demos skip and production deployments live by. What is the agent allowed to do on its own? What needs a human signature? What does failure look like, and where does it log?

Example: parts lookup across supplier catalogues

Surfgreen.dev is building one for an industrial parts distributor in Bavaria. The customer carries catalogues from a long list of manufacturers, each with its own article numbering and lookup quirks. A support employee identifying a replacement used to phone three suppliers and reconcile by hand.

The agent takes a description of the broken part, queries each manufacturer, lines up SKU formats against the customer's own inventory, and returns a candidate with the catalogue evidence. When it is not sure, it says so and shows the alternatives.

It is an agent because it picks the tools and the order itself. Surfgreen.dev supplied the goal and the catalogues. The steps for a hydraulic seal differ from the steps for a control board, and the model works that out.

Where agents fail

Agents fail often. Industry analysts expect more than 40% of agentic projects to be cancelled by the end of 2027. Two reasons keep coming up. The use case did not need an agent in the first place; a workflow would have been more reliable and cheaper. Or nobody fitted guardrails, and the agent did something embarrassing.

A working rule: if the steps fit on a whiteboard and they will still be right next quarter, build a workflow. If the steps depend on what the model finds along the way, an agent is worth the effort.

What this means for German companies

Bitkom's 2026 study counts 41% of German firms actively using AI, more than double 2024. Agentic capabilities sit high on the growth list. 53% name legal uncertainty as the biggest blocker. Seven out of ten German firms have stopped at least one innovation project over data-protection concerns.

That is the gap Surfgreen.dev builds into. An agent that touches your CRM, your ERP, your customer data should do its job without sending the data to a US-hosted model API. We deploy on-premise or on German-hosted infrastructure: Anthropic Claude where the customer holds a Business or Enterprise contract that permits it, and fully local stacks via OpenClaw and NVIDIA NemoClaw where data has to stay on site.

Three questions for evaluation

Three questions tell you what is behind an "AI agent" claim:

1. Who decides the next step — the code or the model?
2. Which tools can it actually call?
3. What happens when it gets something wrong?

If the answers are clear, the term fits. If they are not, what is being described is probably a workflow with a language model bolted in — and very often that is exactly what the use case needs.