agent-keretrendszer

in English: agent framework

advancedspreading

An agent framework is a software toolkit that allows developers to assemble autonomous AI agents without having to write everything from scratch.

In plain terms

Imagine opening a restaurant. You do not build the stove or forge the knives yourself; instead, you rent a fully equipped kitchen where you only need to bring the recipe and the chef. The agent framework is such a ready-made kitchen for creators of AI agents.

A language model by itself only writes text. It becomes an agent by using tools, remembering previous steps and carrying out a task over multiple steps. The framework supplies these recurring parts ready-made: tool integration, memory, step control, error handling and logging.

This is neither the agent itself nor the model working behind it, but the framework between the two. A customer service agent, researcher agent, or coding agent can be built from the same framework, and the model is generally interchangeable underneath.

How it works

The developer specifies the agent's task, the rules it must follow, and which language model serves as its "brain".
It activates the tools the agent can use, such as the search engine, the calendar, a database, or email. The framework passes the description of the tools to the model, i.e., what each is for and how to use it.
During execution, the framework loops the process: the model decides the next step, the framework executes it, returns the result to the model, and this continues until the task is completed.
Meanwhile, it records what has happened so far, manages errors, and enforces limits, such as requiring human approval before payment.
For more complex tasks, it coordinates the work of multiple agents: one researches, another writes, and a third verifies.

An everyday example

In a webshop chat window, you ask where your package is and request rescheduling delivery to Friday. In the background, an agent checks your order, queries the courier's system, and modifies the time. This sequence is likely held together by an agent framework.

Who is affected

developerThey select and use the framework, and this decision largely determines how maintainable and secure the agent will be, and how much work will be required to switch later. → Code

entrepreneur, executiveThe choice of framework determines the cost of deploying an agent, how tied the company is to a single vendor, and who can further develop it later.

employeeAI assistants appearing in your workplace are usually backed by such a framework, so its settings determine what they can access and do on your behalf. → Work & HR

A common misconception

MythIf a company chooses a good agent framework, its agent will automatically become smart and reliable.

In realityThe framework is only a skeleton. The result depends on the quality of the underlying model, well-written instructions, connected tools, and thorough testing.

Limits and risks

The framework does not make the model more reliable: if the model makes a mistake or fabricates something, the agent will act based on that error.
Many convenience layers obscure exactly what happens in the background, making errors harder to find.
The field changes rapidly, so a popular framework today may become obsolete in a year or two, or its updates may require rewriting the code built on it.
The more tools and data an agent has access to, the greater the damage if it receives bad instructions or is intentionally misled.

Similar, but not the same

AI-agent – The agent is the ready-made, task-executing software, while the framework is the toolkit from which it is built.
large language model – The model is the thinking part that generates text, while the framework provides the tools, memory, and control around it.
Model Context Protocol – The MCP standard defines how an agent connects to external tools, while the framework organizes the agent's entire operation and often uses this standard itself.

Related concepts

Broader

szoftver-keretrendszer, AI development tools

Narrower

multi-agent framework, agent SDK

On the edge

AI-agent, coding agent, tool use, orchestration, Model Context Protocol, agent memory

A brief history

The first widely known framework built on language models, LangChain, was released as an open-source project in October 2022. In 2023, Microsoft researchers, together with university partners, published the AutoGen framework, which is based on the collaboration of multiple agents. In 2025, major model development companies successively released their own agent-building kits, including OpenAI, Google, Anthropic, and Microsoft.

Where things stand

By the end of 2026, these tools are established among developers, with many open-source and proprietary versions competing. The Model Context Protocol is widely adopted for connecting tools, but the frameworks themselves are not unified, so an agent written in one cannot be easily transferred to another. Most companies are still building only experimental or narrowly deployed live systems with them.

Status: 2026 November

In the news

Appears in 3 Agent Age stories; the curve shows weekly occurrence in the news magazine.

GitHub Security Lab automates fuzzing with an AI agent frameworkSep 24, 2026 · Code
Critical sandbox escape found in DeepSeek's open-source coding agent Harness (CVE-2026-82533)Sep 8, 2026 · Code
Insurers turn to AI for claims and back-office work: MS&AD, Korea, NTT DATA, GuidewireAug 3, 2026 · Money

Also known as

agent framework · AI agent framework · agentic framework · agent-keretrendszer · agent keretrendszer · agent framework · agent SDK · agent development framework

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.