context-optimization

Optimize context windows by applying compaction, observation masking, KV-cache reuse, and context partitioning. These techniques compress verbose outputs, summarize older context, and split tasks across sub-agents to maintain signal while reducing token costs.

Context partitioning splits tasks across sub-agents to prevent cross-context growth in multi-turn conversations. Use it when processing large documents or managing long-running agent workflows where cumulative context exceeds available budgets.

Reduce token costs by applying compaction to summarize content near limits, observation masking to replace verbose tool outputs with compact references, and KV-cache optimization to reuse computations across requests with stable prefixes.

Compaction summarizes content near context limits to free space while explicitly preserving decisions and context. This ensures critical information survives the compression process when managing long-running agent workflows.

The best approach combines four techniques: compaction for summarization, observation masking for verbose outputs, KV-cache optimization for stable prefixes, and partitioning across sub-agents. A budget-aware workflow governs when to apply each technique.

Observation masking replaces verbose tool outputs with compact references in multi-turn conversations. It reduces token usage while retaining the ability to trace back to original observations when needed within the agent system.