# Context Signal-to-Noise Ratio The context signal-to-noise ratio (context SNR) is the proportion of task-relevant information versus irrelevant information within an AI agent's context window. It's the core metric that [[Context Engineering]] optimizes. ## The optimization target The formal framing of context engineering defines context as **C = A(c_instr, c_know, c_tools, c_mem, c_state, c_query)**, with the goal of maximizing output quality subject to |C| ≤ L_max. In practice, maximizing output quality means maximizing context SNR: ensuring every token in the context window earns its place by contributing to the task at hand. ## What degrades SNR | Degradation type | Mechanism | Note | |---|---|---| | [[Context Bloat]] | Too much context overall | Volume problem | | [[Context Distraction]] | Irrelevant but correct information | Relevance problem | | [[Context Confusion]] | Contradictory information | Consistency problem | | [[Context Poisoning]] | False information | Accuracy problem | | [[AI Context Rot]] | Stale information | Freshness problem | | [[Context Entropy]] | System-level disorder | Structural problem | Each failure mode attacks SNR from a different angle. Effective context management addresses all of them. ## The funnel model The funnel diagram in [[Context Engineering]] visualizes SNR directly: context acts as a filter between all possible answers and the one the model produces. High SNR = tight funnel = specific, accurate answer. Low SNR = leaky funnel = generic or hallucinated answer. ## Improving SNR - **[[Context Budget]]**: hard limits force prioritization - **[[Progressive Disclosure]]**: load context incrementally, only what's needed - **[[Context Layering]]**: separate concerns so irrelevant layers aren't loaded - **[[Context Lifecycle]]**: regular review and pruning remove noise - **[[Receptionist AI Design Pattern]]**: route to specialized agents with focused context - **Retrieval precision**: tune RAG for precision over recall ## References - ## Related - [[Context Engineering]] - [[Context Budget]] - [[Context Bloat]] - [[Context Distraction]] - [[Context Confusion]] - [[Context Poisoning]] - [[AI Context Rot]] - [[Context Entropy]] - [[AI context is finite with diminishing returns]] - [[Context Window]] - [[Progressive Disclosure]] - [[Context Layering]] - [[Context Lifecycle]] - [[Receptionist AI Design Pattern]] - [[Natural tension between compression and context]]