The Noise Floor Problem in AI Agent Context Windows
The Noise Floor in Agent Context Windows
Think of context windows like a radio signal. Every irrelevant piece of information you add raises the noise floor. Once the noise floor gets high enough, the model cannot reliably extract the signal - even if the right information is technically present in the context.
What Noise Looks Like in Practice
- Full stack traces when the agent only needs the error message
- Complete file contents when only 3 lines are relevant
- Verbose tool outputs that repeat information the agent already has
- Historical conversation turns that no longer matter for the current decision
- System prompts that describe capabilities the agent will never use in this session
Each of these individually seems harmless. Together, they push the context toward a state where the model starts making worse decisions - not because it lacks information, but because the right information is buried.
Measuring Your Noise Floor
A simple test: take your agent's context at any decision point and ask - what percentage of these tokens are directly relevant to the next action? If the answer is below 30%, your noise floor is too high.
Practical Noise Reduction
Summarize tool outputs before adding them to context. If a tool returns 2,000 tokens but the relevant information fits in 50 tokens, summarize before injecting.
Use sliding windows for conversation history. Keep the last 3 to 5 turns in full detail, compress everything else into a running summary.
Scope system prompts to the current task. An agent doing file organization does not need instructions about browser automation in its system prompt.
Strip formatting and metadata. JSON responses often include metadata fields that the agent never reads. Remove them before they enter the context.
The Payoff
Clean context does not just save tokens - it improves decision quality. Agents with focused, signal-rich context windows make fewer errors, recover faster from mistakes, and complete tasks in fewer steps.
Fazm is an open source macOS AI agent. Open source on GitHub.