What is an AI agent?

An AI agent is a language model that works through a task with tools until the task is done. It reads the input, chooses a step, calls a tool and looks at the result. Then it chooses the next step. An agent can read an order confirmation from an email and enter the delivery date in the customer record of a CRM.

How an AI agent differs from a chatbot and an LLM

A language model (LLM) produces text from text. A chatbot puts that model behind a chat window and answers questions. An agent uses the same model and acts in other systems.

Anthropic draws a second line in its guide Building effective agents. In a workflow, code fixes the path and the model handles single steps. In an agent, the model directs its own steps and its use of tools.

What an AI agent can do, and when a script is enough

An agent is useful where the input varies: emails in free text, scans, PDFs in different layouts. These are tasks where a person reads and decides today.

Many routine tasks need no AI. A script fits when the format stays the same. A workflow tool fits when a fixed path runs across several systems. Before a company picks one, I compare the three options by input, rule, cost per run and typical failure.

An agent reaches a system through an interface, an export or a database it may read. The Model Context Protocol (MCP) is an open standard for that access. An MCP server gives an agent access to one of your systems, with the rights you grant it.

What is an agent harness?

An agent harness is the program around the model. It hands the task to the model, runs the tools the model calls and loads the rule files. Claude and Codex each have their own harness, and each reads its instructions from a different file.

What you decide before an agent takes over a task

An agent follows a written rule most of the time. A fixed program stops at an unknown format. An agent goes on with a wrong reading. A written rule alone therefore does not bind it, and four limits do the work:

• Rights at the system. The agent’s database role reads and cannot write. Its keys belong to the test system.
• A test system first. The agent does the task there while your staff do it as before.
• A confirmation before a change to live data. A person gives it outside the chat, and it covers one change.
• A read-only script that measures the result. When the script and the agent disagree, the script’s number stands and the agent runs again.

Two decisions stay with you. You name the person who confirms changes to live data. You decide which systems the agent may read and which it may change.

My work with AI agents

I am an IT expert for data platforms. For more than twenty years I have connected systems, from the payment connections of online shops to an existing CRM. Today AI agents speed up how I build and transform data platforms.

I help a client rebuild its platform with Claude and Codex as coding agents. On this project the AI agents work under the same rules as humans. I have set the limits at the resource and have written and tested the blocking checks.

For the same client I set up agents with roles on a control plane. My research on virtual organizations of autonomous agents asks how a company works when most of its staff are software agents. At Claude Hacker House I teach building MCP servers. I take on agent projects as a freelancer or as a permanent employee.