Long-Horizon

Some tasks are done in a single step. Translating a word or solving an arithmetic problem falls into this category. Other tasks require many steps that build on one another. Moving house, for example: getting boxes, packing, renting a van, changing your address, unpacking. In computer science, tasks with such a long time horizon are called long-horizon. The English term literally means “long horizon” and refers to the fact that the goal lies far away, with many decisions in between.

Why errors snowball over many steps

The key point is simple math. Suppose a program completes each individual step with 95 percent reliability. That sounds very good. But over twenty consecutive steps, the chance of an error-free run drops to only about 36 percent. After a hundred steps, it is practically zero. This is exactly where AI systems still often fail today.

On top of that comes a second problem: errors don’t simply sit harmlessly side by side. A wrong step changes the starting conditions for every step that follows. Anyone who books the moving van for the wrong day can’t get anything else right afterward. In technical terms, this is called error accumulation. A system that doesn’t notice and correct its own mistakes drifts further and further away from the goal.

Economically, this makes a huge difference. A chatbot that answers questions is useful, but the human still checks every answer themselves. A system that independently carries out a travel booking or a software change from start to finish, on the other hand, saves real working hours. That’s why long-horizon capability is currently considered one of the most important hurdles in AI development.

Planning, remembering, verifying

The first building block is planning. The system breaks the big goal down into smaller sub-goals before it starts acting. So instead of acting immediately, it writes itself a kind of to-do list. It can adjust this list later if something doesn’t work out.

The second building block is memory. A language model can only take in a limited amount of text at once; this window is called the context. For long tasks, everything eventually no longer fits. That’s why modern systems store notes in files or databases and later retrieve exactly what they need. In effect, they keep a kind of project diary.

The third building block is feedback from the environment. If a system writes program code, it can execute it and read the error message. This way, it notices after every step whether it’s still on track. Tasks with this kind of automatic feedback succeed considerably more often than tasks without it. Incidentally, a common misconception is that a bigger model automatically solves this problem. More knowledge doesn’t help much if the structure for self-checking is missing.

From agent demos to benchmark tables

The term appears most often in connection with AI agents. These are programs that operate tools independently, such as a browser or a development environment. When a provider promises that its agent can work independently for hours on end, they mean precisely this long-horizon capability.

In technical articles and quarterly reports, the term also appears as a metric. There are benchmarks that indicate how long a task would take a human to complete, which the system can still reliably handle. These figures have risen from a few minutes to several hours in recent years. In practice, such numbers often form the basis for market expectations.

The term originally comes from reinforcement learning, where a program learns through rewards. There, the problem was that a reward only arrives right at the end, for instance upon winning a game. Today, long-horizon is used more broadly, including for robots and language models. Anyone who understands the basic idea will immediately grasp, when reading product announcements, where the real progress lies.