The Cure for Context Rot

Context rot is the slow death of an AI agent’s attention span and one of the key problems holding agents back. Essentially, as an agent accumulates more information their performance degrades, sometimes to the point of failure.

The term was coined in June 2025 by Hacker News user Workaccount2 and caught on quickly as agent developers recognized the behavior from their own work. Just a few weeks later, “context rot” is showing up in Twitter threads, Linkedin posts, and many more Hacker News comments. Clearly, it’s a problem that resonates.

We built Subconscious to fix it. With our inference engine (a reasoning model and inference runtime designed to run together), you can run agents that stay as focused at token 100,000 as they were at token 1,000. Let’s break down how we approach the problem.

How reasoning models work today

Most frontier models are trained to reason step by step. They process tasks in a linear chain of thought, carrying forward everything they’ve seen so far. That works fine for simple workflows, but breaks down under the complexity of long running tasks. If you've ever asked ChatGPT a long series of questions or asked cursor for 10+ revisions across multiple files, you've felt the performance degradation.

Take the search query: “Tell me about three early stage startups that might be the next Figma.” To answer this, the agent has to understand the prompt, search for candidate companies, interpret the results, and possibly search again to verify or compare.

The beginning of that process might look like this:

Example of a linear chain of thought agent with a search tool.

Each tool call adds a large chunk of data to the context window, in this example search results. This works fine for short tasks, but as the chain grows, the context starts to bloat. This problem compounds if the agent takes a wrong turn, like exploring Adobe Firefly (not a startup), as it adds noise without getting closer to a valid answer. The more steps and detours, the more noisy the context becomes. That’s when context rot sets in.

Learning more about Adobe Firefly does not improve the search results, but it does add more tokens to the context window.

As the reasoning chain grows, performance drops, and the agent becomes more likely to produce a poor quality result. For engineers and their end users, that means wasted time, tokens, and effort. It’s a frustrating process.

Teams can mitigate rot with context engineering. One option is multi-agent systems that break the task into smaller pieces for subagents to handle. Another is pruning the context between steps to keep relevant information that matters. Both approaches can work, but they add engineering complexity and overhead. Walden Yan at Cognition wrote about building single-agent systems that manage their own context, and we agree with him that single agent systems are the better approach.

How Subconscious Solves Context Rot

We built Subconscious because we ran into context rot ourselves while trying to build production grade agents. The core of our system is a model and runtime designed to work together: TIM and TIMRUN. TIM handles task decomposition and tool use. TIMRUN understands the structure of TIM’s output and prunes irrelevant context between steps. Together they expertly break down tasks and extend agent usefulness beyond their context limits.

Let’s look at the same “next Figma” query. TIM starts by generating a list of promising candidates. For each one, TIMRUN spins off a subtask to explore it in more depth.

A reasoning tree forms to explore each candidate company

After each subtask gathers information using the search tool, TIMRUN passes only the relevant company data back to the parent thread.

Once the branches have been explored, relevant information is passed to the parent node.

At that point, the model prunes everything it no longer needs and moves forward with a clean context. After extensive search and processing many tool tokens, the TIM model still perfectly follows the original user instruction because the attention head is still in near proximity from the original input after pruning.

As the reasoning tree continues, child nodes are automatically removed from the context window by TIMRUN.

Accurate long-horizon reasoning is only possible when you control the context.

Context is the answer

Context rot is a key obstacle keeping agents from going into production. Investing weeks in context engineering for each new feature could work, but we built Subconscious to solve the problem for you automatically.

If you want to talk to us about how we're solving context rot, schedule a call to chat with us for 15 minutes.