context-optimization

Context Optimization Techniques

Context optimization extends effective capacity through strategic compression, masking, caching, and partitioning. The goal is making better use of available capacity — not magically increasing windows. Effective optimization can double or triple effective context without larger models.

When to Use / Not Use

Use when:

• Context limits constrain task complexity
• Optimizing for cost reduction (fewer tokens = lower costs)
• Reducing latency for long conversations
• Implementing long-running agent systems
• Building production systems at scale

Do NOT use when:

• Compressing or summarizing context -> use context-compression
• Diagnosing context failures or degradation -> use context-degradation
• Learning context basics -> use context-fundamentals
• File-based context patterns or scratch pads -> use filesystem-context

Decision Tree

What optimization problem are you solving?
├── Tool outputs dominate token usage (>80%)
│   └── Observation Masking -> Replace verbose outputs with references
├── Context approaching limits (>70% utilization)
│   ├── Message history dominates -> Compaction + Summarization
│   ├── Retrieved docs dominate -> Summarization or Partitioning
│   └── Multiple components -> Combine strategies
├── Repeated requests with common prefixes
│   └── KV-Cache Optimization -> Stable prefix ordering
├── Single context too large for one agent
│   └── Context Partitioning -> Sub-agent isolation
├── Need to measure if optimization is working
│   └── See §Performance Targets
└── Not about extending capacity? -> See related skills

Four Primary Strategies

Strategy	Mechanism	Best For	Expected Savings
Compaction	Summarize context near limits, reinitialize with summary	Message history dominating	50-70% token reduction
Observation Masking	Replace verbose tool outputs with compact references	Tool output dominance (80%+ of tokens)	60-80% reduction in masked obs
KV-Cache Optimization	Reuse cached KV computations across shared prefixes	Repeated requests with stable prefixes	70%+ cache hit rate
Context Partitioning	Split work across sub-agents with isolated contexts	Single context too large	Isolation + clean focus