Long-Running Agent
A long-running agent is an AI program that autonomously pursues a task over hours or days, rather than simply answering a single question. It plans intermediate steps, uses tools such as search or program code, and keeps track of its progress so far.
Most artificial-intelligence chat programs operate in a very short cycle. You ask a question, the program answers, and the process is over. A long-running agent works differently. It is given a larger assignment and keeps working on it for hours or even days. In doing so, it breaks the task down into small steps itself and completes them one after another. In between, it checks its own intermediate results and changes its plan if something goes wrong. So the human no longer dictates every single action, but only sets the goal.
From answer-giver to contractor
The difference sounds technical, but it fundamentally changes how the technology is used. A short answer can be read and judged in ten seconds. An eight-hour assignment can no longer be checked so easily. This is exactly where the appeal lies: tasks that were previously too big for an AI become possible. One example is rewriting an old program into a new programming language.
Economically, this is the reason why many companies are currently pouring so much money into this direction. A tool that answers questions saves minutes. A system that independently completes entire work packages replaces hours of labor. Accordingly, expectations are high for companies such as OpenAI, Anthropic, or Google, which offer such agents.
At the same time, the risk grows. An error in the third step can carry through all the following steps. If no one is watching, hours of unusable work can pile up. That is why developers build in checkpoints where a human must confirm before the process continues.
Plan, tools, and memory
At its core, an agent consists of three parts. First, a language model — the program that understands and generates text. Second, tools: this can be an internet search, a program for executing code, or access to files. Third, a loop that keeps asking: What is the next step, and am I already done?
The biggest technical problem is remembering. A language model can only take in a limited amount of text at once. After many hours of work, the entire history no longer fits. Agents solve this by writing notes into files and summarizing older sections. You can picture this like a lab notebook in which only the most important results end up.
A common misconception is that a new kind of AI is at work here. The language model on the inside is usually the same one used in ordinary chat. What is new is only the environment surrounding it: the loop, the tools, and the storing of intermediate states. Experts therefore speak of a scaffold, or scaffolding.
Agents in products and headlines
So far, long-running agents are most widespread in programming. Tools such as Claude Code, OpenAI Codex, or Google’s Jules are given a task along the lines of “fix this bug” and then work alone on a project. They read files, change code, run tests, and repeat this until the tests pass. In the end, they submit their change for review.
Outside the software world, similar systems appear in research. So-called deep-research features search through hundreds of web pages over an extended period and write a report from them. Agents that independently operate a browser also belong to this category. They fill out forms or compare prices without anyone clicking along.
In the news, the term usually appears in two contexts. Either it is about records, for instance when a provider announces that its agent worked for thirty hours without interruption. Or it is about safety and liability, because an agent with far-reaching permissions can also cause damage. Both debates ultimately revolve around the same question: How much autonomy do you give a program?