Focus: attention isn’t enough

​Attention - noun

At·​ten·​tion ə-ˈten(t)-shən
The act or state of applying the mind to something

Focus - noun

fo·​cus fō-kəs
A center of activity, attraction, or attention
a point of concentration
directed attention : emphasis

AI coding assistants have transformed programming, but they share a common fault with users – they forget. As tasks progress and prompts grow, these agents gradually lose sight of earlier details. Even as the context windows of frontier models expand, it’s tempting to assume we can await the sunset of this problem. Unfortunately, simply increasing context size hasn’t eliminated it. In fact, once a large language model’s context window begins to fill up, the quality of its outputs often decreases. This “lost in the middle” phenomenon has been documented by researchers: models tend to focus best on information at the beginning or end of their input and significantly degrade when important details are buried in the middle. In other words, LLMs are as susceptible to information overload just as we are.

Bigger windows, same memory woes

During extended multi-step tasks, LLM agents often experience goal drift, a phenomenon rooted in physical context window constraints. As a model generates tokens, earlier steps may fall outside the active attention window—GPT‑family models, for instance, exhibit the Lost in the Middle effect — a U-shaped performance curve where information at the start or end of the context is recalled far better than details buried in the middle. Without explicit anchoring mechanisms, the agent simply cannot “see” its own prior rationale, increasing the risk of re-executing actions, contradicting previous decisions, or drifting off task.

Efforts to extend LLM context windows (now up to 1M+ tokens) are impressive, but longer context alone isn’t a silver bullet. Evaluations have shown that many models’ performance saturates—and often degrades—beyond certain thresholds. This is not a new problem: Greg Kamradt observed that GPT-4’s accuracy noticeably declined when given more than 64K tokens of context out of its full 128K window. Context management strategies remain crucial, even with today’s increased limits. Researchers generally pursue two complementary approaches: (i) build models that can technically handle longer contexts, and (ii) use retrieval or summarization to supply only the most relevant snippets (arxiv.org).

While models are increasingly able to digest large amounts of input, there remains a sweet spot in the context window where outputs are at their sharpest. As IBM’s AI team said long ago (in AI years), “like people, LLMs are susceptible to information overload… throw too much detail at them, and they may miss the key takeaways.” Even today with larger token windows on the horizon, well-structured context matters.

We’ve learned a bitter but essential truth: scale alone is not enough—general methods perform best when provided with the right information in the right form. For Cline, this means resisting the reflex to rely on RAG. Retrieval is passive; agency is active. Segmenting a repository into disconnected, vector-indexed chunks and relying on cosine similarity to surface the right ones risks diluting reasoning and distracting the model. A capable agent should explore the codebase like a senior engineer—navigating folders, following imports, reading files in sequence, and building a coherent internal map of the problem space. That is why, instead of indiscriminately dumping the entire codebase or depending on brittle RAG pipelines, we supply the model with a clear, scoped “to-do list” at each user turn: task-aware, narrative-driven context that is aligned, scoped, and relevant to the task at hand. Everything it needs, and nothing it doesn’t.

Not all tokens in a context window are created equal. Tokens that directly advance the task — key requirements, critical code paths, or decisions the agent must act on, can be seen as high-value tokens. Low-value tokens are noise: redundant code, tangential comments, or repeated snippets that consume space without improving reasoning. In large codebases, it’s easy for low-value tokens to crowd out the high-value ones, especially when relying on broad retrieval pipelines. The result is a bloated prompt that costs more, runs slower, and often performs worse. We aim to emphasize high-value tokens: where every step in a plan is deliberate, relevant, and positioned where the model can use it effectively. By anchoring the agent on what matters most, we make better use of every token — and focus the model’s attention.

A context-forward approach to staying on target

We’re rolling out a new feature to tackle this problem head-on: the Focus Chain. This is not just a to-do list or a UI nicety – it’s a context-forward approach to single-task orchestration, purpose-built to keep an AI agent’s attention anchored on the task at hand. In essence, the Focus Chain has the agent generate a step-by-step plan for the task, and persistently carry that context forward as it works. Each step of the way, the agent refers back to and updates the plan, ensuring it doesn’t lose sight of what it’s doing and why.

Here’s how it works in practice: when you give Cline a request, the very first thing it does is draft a numbered list of all the steps it will need to complete (this is the “chain” of focus). This plan is kept in the agent’s working context. At every subsequent stage, the agent checks the Focus Chain, marks progress, and revises the remaining steps if necessary. Crucially, the same model is tasked with updating the list over time. By doing so, the plan itself becomes part of the prompt, reminding the model of past actions and upcoming actions. The agent essentially “remembers what it is working on” because the evolving to-do list travels with it through the context. This continuous grounding in the original design prevents the model from straying off course or chasing tangents. It’s a bit like an AI project manager whispering in the agent’s ear: “Stay on target – here’s what we’re doing.”

Staying focused in this way leads to measurably better results, especially for complex, long-running tasks that previously tended to go off the rails. By preserving the chain-of-thought, Focus Chain mitigates the pseudo-amnesia that can plague lengthy LLM sessions. The agent is far less likely to repeat earlier steps or contradict itself, because it isn’t relying purely on an ever-fading latent memory – the explicit plan is always in view.

Other coding agents have released similar features, but ours takes a context-forward approach that we believe makes a meaningful difference. This initial release is a strong foundation, delivering value now while paving the way for even greater capabilities ahead.

Adding links to the chain

To demonstrate the impact of the Focus Chain, we’re also releasing a new workflow, called “deep-planning”, designed to demonstrate the Focus Chain’s capabilities. In this workflow, the AI agent is first tasked with exploring your codebase in depth and producing a comprehensive implementation plan for a requested feature or fix. This is effectively a planning phase: the agent reads relevant files, gathers context, and uses that information to draft a detailed plan of attack. Once the plan is ready (and vetted), the workflow spins up a second task – the execution phase – where Cline actually implements the changes. Here’s where the Focus Chain shines: the plan created in the first phase is carried over into the second phase as an always-present guide. The agent tackling the implementation has all the details it needs from the planning phase, and it leverages the Focus Chain to remain laser-focused on the user’s request throughout the potentially long coding session. Even if the task ends up consuming millions of tokens of generation (think hours of work equivalent), the agent persists, the Focus Chain keeps it on what it was asked to do.

This two-stage workflow is just one example of how the Focus Chain can be used to boost efficiency and reliability. By splitting a complex job into a planning stage and an execution stage – and maintaining a through-line of context via the chain – we set the agent up for success. It approaches the coding task with a clear roadmap and stays on track, no matter how large the codebase or how lengthy the session. We believe our community will discover many other creative uses for the Focus Chain. Whenever you have a complex task that could benefit from a persistent, structured context, the Focus Chain might be the key to keep your AI on task and on target.

Concluding this context

As AI capabilities advance, it’s important to address not just how much these models can process, but how well they utilize what they’re given. Larger context windows will help, but they won’t completely sunset the memory limitations of LLMs on their own. Focus Chain is our answer to this challenge – a way to enrich an AI agent’s short-term memory and guard against the natural tendency to forget. By moving context forward and orchestrating tasks step by step, we’re aiming for consistency, accuracy, and focus, even on the longest and most complex coding tasks.

The early results are exciting: long-horizon tasks that used to derail partway through now stay coherent to the end, and agents produce code more aligned with the initial design. We’re eager for you to try Focus Chain in your own projects and workflows. This feature represents a new level of reliability for AI coding assistants, and we’re just beginning to explore its potential. With the Focus Chain keeping our models on track, we can push the boundaries of what’s possible – without forgetting where we started.