context-optimization

Reduce token usage in long-running agent workflows by applying observation masking, context compaction, and KV-cache optimization. These techniques preserve task state while cutting token costs within limited context windows.

Observation masking reduces token usage by replacing verbose tool outputs with compact references. This technique preserves critical state information while cutting the tokens consumed by long histories.

Summarize long histories without losing task state by using context compaction. This technique preserves critical state information while significantly cutting the tokens consumed by long histories.

Yes, you can split large documents across sub-agents using context partitioning. This enables scalable reasoning while maintaining a modular and debuggable design for large problems.

Track token usage and trigger context optimizations automatically by using budgeting and monitoring hooks. These hooks monitor consumption and invoke optimization passes when context usage approaches limits.

Avoid context partitioning when tasks require tightly coupled state or cannot be split into independent sub-problems. It is best suited for large documents and workflows requiring scalable reasoning.