Glossary › Agents

long-horizon task

A complex task that an AI agent can only perform by working independently over a long period through many sequential steps.

In plain terms

Answering a question is like telling someone where the nearest hardware store is. A long-horizon task is more like renovating a bathroom: you must assess, plan, buy materials, demolish, tile, and keep the original goal in mind throughout.

In AI, this is what we call tasks that consist of dozens or even hundreds of steps and would take a human hours or days. Examples include writing and debugging a program, conducting market research, or handling a case that involves multiple systems.

The difficulty here lies in persistence and accuracy, because individual steps can be easy on their own. If an agent solves each step with 99 percent certainty and does not notice its errors, after a hundred steps the chance of making no mistakes at all is only about 37 percent, or just over one-third.

This is why it has become one of the most important metrics for agents: what matters is how long a task can be entrusted to the model without supervision, and answering individual questions well is not enough for this.

How it works

1. The agent breaks down the received goal into subtasks, sets an order, and thus creates a plan.
2. It works step by step: it uses tools (searches, reads a file, runs a program), then checks the result and decides on the next step based on it.
3. Since the model can only keep a limited amount of text in mind at once, it keeps notes, summaries, or external memory about what it has already done and what remains.
4. It occasionally checks its own work (for example, by running tests), and if it finds an error, it steps back and fixes it to prevent the error from compounding.
5. For larger tasks, the work is distributed among multiple agents, and a human approves continuation at certain points.

An everyday example

Who is affected

A common misconception

Limits and risks

• Errors accumulate: an agent may build on an early bad decision for hours before the problem is revealed, so intermediate checkpoints are advisable.
• The measured 'time horizon' typically refers to the task duration (calculated in human expert working time) at which the model solves roughly half of the tasks, so it is not a guarantee of success. Moreover, measurements are mainly conducted on software tasks, so they can only be cautiously transferred to other fields.
• A long run consumes a lot of computing capacity, so costs are hard to estimate in advance, and a stuck agent can unnecessarily burn money.
• The longer an agent works without supervision, the more damage a mistaken action can cause, so require human approval for irreversible steps (payment, deletion, sending).

Similar, but not the same

• context window – The context window shows how much text the model sees at once, while the long horizon indicates how long a task it can complete; the latter depends on many factors beyond just the window size.
• multi-step reasoning – Multi-step reasoning is a chain of thought within a single response, whereas long-horizon tasks involve a long sequence of real actions using tools, feedback, and corrections.
• workflow automation – An automated workflow executes pre-written, fixed steps, whereas in a long-horizon task, the agent decides the next step on the fly.

Related concepts

A brief history

The term 'horizon' originates from reinforcement learning and planning theory, where it denotes how many steps ahead to calculate. It became a widely recognized metric for large language models in 2025, when the METR research organization published a study measuring how long tasks AI models can perform in human work hours. It found that this duration roughly doubled every seven months between 2019 and 2025; later measurements by the organization suggest growth has been even faster since 2024.

Where things stand

At the end of 2026, the best models can complete even well-defined tasks equivalent to many hours of human work, especially in software development; measurements suggest this length may already reach or exceed one workday, but the estimates are highly uncertain. Reliability still noticeably declines as task length grows, so work lasting days with no supervision at all remains an exception for now.

Status: 2026 November

In the news

• FIS launches agentic AI suite for CFOs with cash flow forecastingOct 12, 2026 · Money

Also known as

long-horizon task · long horizon task · long-horizon task · long-horizon task · long-running task · long-running task · time horizon · long-horizon agent

Sources

• Kwa et al. (METR): Measuring AI Ability to Complete Long Tasks (arXiv, 2025)
• METR: Measuring AI Ability to Complete Long Tasks (summary)

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