Context Optimization

What it does

This skill, Context Optimization, helps AI agents work effectively even with limited context windows. It achieves this by strategically compressing, masking, caching, and partitioning conversation history. The goal is to improve response accuracy and relevance without needing larger models or longer context windows – ultimately increasing efficiency and reducing costs.

When to use it

• When task complexity is constrained by context limits.
• To reduce the cost of running an AI agent (fewer tokens used).
• For long-running agent systems where latency needs to be minimized.
• When handling larger documents or conversations.
• When building production systems at scale.

Key capabilities

• KV-cache optimization: Reorders and stabilizes prompts for improved inference engine efficiency.
• Observation masking: Replaces verbose tool outputs with compact references, preserving retrievability.
• Compaction: Summarizes context when utilization exceeds 70%, prioritizing compression of tool outputs, old conversation turns, and retrieved documents (excluding the system prompt).
• Context partitioning: Splits work across sub-agents with isolated contexts for tasks exceeding a certain size threshold.

Example prompts

• "Optimize the current conversation history to reduce token usage."
• "Apply context masking techniques to shorten this exchange."
• "Can you summarize the key points from our previous discussion?"

Tips & gotchas

• Always apply KV-cache optimization first, unconditionally.
• Context quality is more important than quantity; measure before and after optimizations.
• Avoid compressing the system prompt as it anchors model behavior.