Context Window Pollution
A phenomenon occurring during long-chain task execution by AI Agents, in which the context window becomes filled with large volumes of low-value information such as tool call logs and intermediate outputs. This reduces the density of meaningful information and degrades the quality of the model's reasoning.
Timeline (last 90 days)
As the number of agent execution steps increases, a large accumulation of tool-call records and intermediate outputs in the context causes the density of useful information to decline and affects model decision-making; this is referred to as context window pollution.
Context window contamination refers to the model taking the entire conversation history as input, where early content continuously occupies the token quota and may interfere with the model's judgment on the current task.
All Facts (2)
As the number of agent execution steps increases, a large accumulation of tool-call records and intermediate outputs in the context causes the density of useful information to decline and affects model decision-making; this is referred to as context window pollution.
50%UnverifiedContext window contamination refers to the model taking the entire conversation history as input, where early content continuously occupies the token quota and may interfere with the model's judgment on the current task.
50%