What is an AI agent?

How an AI agent works

At the core of every agent is a large language model. It receives a goal, instructions and a set of tools. In a loop, the model decides which tool to call next, receives the result and plans the next step. Typical building blocks are:

• Instructions: role, tone, rules and boundaries of the agent.
• Knowledge: access to documents, for example via RAG.
• Tools: actions such as searching, reading calendars, drafting e-mails or creating files – often connected via the Model Context Protocol.
• Triggers: a chat message, a schedule or an event that starts the agent.
• Memory and history: what happened in earlier steps and runs.

Chatbot, automation, agent – the difference

A chatbot answers questions. A classic automation follows a fixed if-then flow. An agent combines both: it can react flexibly to unforeseen situations, but works within defined tools and permissions. In practice, the best results often come from a mix – fixed steps where reliability counts, model decisions where language and judgement are needed.

Where agents help in the Mittelstand

• Preparing offers from enquiries and existing price documents
• Daily summaries of e-mails, appointments and open tasks
• Research and pre-qualification of leads
• Answering internal questions on processes and guidelines
• Preparing reports from several data sources

Risks and how to control them

• Wrong actions: an agent can misinterpret a task. Risky actions such as sending, deleting or paying should require confirmation.
• Excessive permissions: an agent should only see and do what the person or role behind it is allowed to.
• Hallucinations: results must remain verifiable – with sources and traceable steps, see hallucination.
• Traceability: a run history shows which steps and tools were used.
• Regulation: depending on the use case, transparency and documentation duties under the EU AI Act may apply.

AI agents in AIPE

In AIPE, you create your own assistants with instructions, documents and tools. In the Agent Studio, automations are built as flows, tested, published and tracked via a run history. Routines run prompts on a schedule or once; runbooks and Claude-compatible skills can be added. Agents use enabled connectors such as Gmail, Google Calendar, Google Drive, Nextcloud, HubSpot or ClickUp as well as configured MCP servers, each with the permissions of the organisation. With the AIPE Worker, tasks can be carried out on your own computers – in "Ask first" mode, you approve risky actions. More in the guide AI agents and on the agents page.