Agent Framework
An agent framework is a ready-made software foundation that developers use to build programs that carry out tasks independently in multiple steps. It provides the building blocks for this: planning, memory, access to tools such as search engines, and the repetition of steps until the goal is reached.
A language model like the one behind ChatGPT is very good at generating text. But on its own it cannot look things up on the internet, save a file, or carry a task through multiple steps. This is exactly the gap an agent framework fills. It is a collection of ready-made program building blocks that developers use to embed a language model into what is called an agent: a program that takes on a goal and independently handles intermediate steps. In doing so, the agent is allowed to use tools, meaning it can call other programs, such as a search engine or a calculator. Well-known examples of such frameworks are called LangChain, LlamaIndex, CrewAI, or AutoGen.
Why nobody builds the loop themselves
One could also program an agent entirely by hand. In practice, however, the same structure keeps repeating itself. The agent has to read out the model’s answer, recognize which tool is meant, call it, return the result, and start the round again. On top of that come unpleasant details: crashed tools, overlong texts, infinite loops. A framework has already handled these cases.
The second reason is interchangeability. Frameworks communicate with the models of various providers through the same interface. Anyone who uses a model from OpenAI today and one from Anthropic or a self-hosted one tomorrow then only needs to change a few lines. This is economically important, because the prices and quality of the models shift on a monthly basis.
Third, companies need traceability. An agent makes many small decisions, and if the result is wrong, someone wants to know why. Frameworks therefore log every step. This record is called a trace, and without it an agent is barely maintainable in production.
Planning, calling tools, remembering
The core of every framework is a loop. The model receives the task and describes its next step. The framework executes this step and sends the result back into the model. This repeats until the model says it is done, or until a fixed maximum number of rounds is reached. This upper limit is not a minor detail—it protects against agents that go around in circles while burning money.
For the tools, the developer stores a short description in everyday language for each one. For example: “query_weather, expects a place name, returns the temperature.” The model chooses on its own based on these descriptions. So it is not the programmer who decides in advance which tool is used when. That’s why a good tool description is often more important for quality than the framework itself.
On top of this comes a memory. Language models forget everything that is not in the current text, because their working memory per request is limited. The framework therefore manages what is sent along and what is offloaded and retrieved later. Some frameworks also let several agents work together, for example one that researches and one that checks. The framework should not be confused with the model: the framework does not think, it only organizes.
From coding assistance to customer service
Today, agents are most visible in programming. Tools such as GitHub Copilot Workspace, Cursor, or Claude Code read a project, change several files, and run tests until they pass. That is exactly an agent at work: setting a goal, taking steps, checking the result, refining.
In companies, agents mostly work invisibly in the background. A customer service agent reads the email, looks up the order number in the database, checks the delivery deadline, and drafts a reply. A research agent gathers figures from several reports. Such systems are almost never built from scratch, but on top of a framework.
In business news, you often encounter the term in connection with money and market power. By now, almost every major provider has its own agent framework, because that is where it is decided which model companies will use later on. Caution is warranted with promises: a framework does not make agents reliable. If the model misjudges an intermediate step, it builds further on top of that error. That’s why, in practice, a human is usually placed as a control before critical actions such as payments.
