Как ИИ-агенты остаются в контексте длинных задач

When you're working on a large project in a code editor, tasks are rarely completed in just a few steps. You often have to go through several files, adjust the architecture, find a bug, and then return to where you started. It's a long process – and this is precisely where AI assistants traditionally run into problems.

The Cursor team encountered this issue while developing their Composer agent and found a way to get around it – not by increasing the model's memory, but by changing how it's trained.

Контекстное окно языковых моделей и их ограничения

The Context Window – Not Just a Technical Limit

Every language model has what's called a «context window» – roughly speaking, the amount of information it can hold «in its mind» at one time. Imagine reading a book but only being able to keep the last 50 pages in your head. Everything that came before no longer affects your conclusions.

For short tasks, this isn't critical. But when an agent is working on something complex – a long chain of actions, a major refactoring, multi-stage debugging – the useful history quickly exceeds this window's limits. And the model starts to «forget» earlier steps, lose the thread of the task, and repeat itself.

One common way to combat this is summarization: the agent periodically compresses its action history into a brief summary to free up space for new steps without losing the essence of what has already been done. Composer does exactly this – it can condense previous actions into a compact description and continue working from there.

Sounds reasonable. But this is where a non-obvious problem arises: how does the model know how to summarize correctly?

Ограничения обучения ИИ-амоделей на коротких траекториях

When a Model's Training Doesn't Match Its Task

Agents are usually trained on trajectories – sequences of actions that have led to a correct result. But the length of these trajectories is limited by that same context window. Anything that falls outside its scope simply isn't included in the training.

This creates a vicious cycle: the model learns to work with short histories but in practice has to deal with long ones. In such a setup, summarization exists on its own – its quality is in no way linked to how well the agent ultimately performs the task.

Simply put: a model can learn to create elegant summaries that, in reality, don't help it see the task through to completion.

Интеграция суммаризации в процесс обучения ИИ-моделей

The Idea: Make Summarization Part of the Training

The Cursor team proposed a different approach. Instead of treating summarization as an auxiliary function, they built it directly into Composer's training process.

Here's how it works conceptually. You take a long trajectory – one that doesn't entirely fit within the context window. It gets broken down into parts. The first part is «condensed» into a summary. Then, the model is trained to continue the task based on that summary – and so on down the line.

The key insight is this: the quality of the summarization is now evaluated not in the abstract, but by how well the model handles the rest of the task. If a summary is inaccurate or omits important details – the model will stumble on subsequent steps. This itself is the training signal.

Thus, the model learns not just to retell its action history, but to do so in a way that the retelling is actually useful for continuing the work.

Важность решения проблемы контекста для ИИ-агентов

Why This Is Important Right Now

As AI agents take on increasingly complex tasks, the ability to work with long horizons is becoming crucial. An agent that « forgets» the context after 20 steps simply won't be able to handle a serious engineering task.

Increasing the context window is one path, but it's expensive and has practical limitations. Teaching a model to efficiently compress and use its own history is another path and, by all appearances, a more flexible one.

Cursor's approach is interesting because it doesn't require a fundamentally new architecture. It's more of a change in training logic: summarization ceases to be an «add-on» and becomes part of what the model is graded on.

Будущие перспективы метода суммаризации для ИИ-агентов

What Remains Unclear

For now, this is an approach described in a company blog post, not an independently verified result. How well it works in practice – outside of test conditions, on real projects with unpredictable structures – remains to be seen.

It's also an open question how well the model handles very long chains – when there are multiple summaries, and each subsequent one builds on the last. Errors here could accumulate.

Nevertheless, the direction seems logical. Agents that know how to «remember» correctly will be able to take on more complex tasks – and see them through to the end.