context-optimization

Context optimization extends effective capacity through compaction, masking, KV-cache optimization, and partitioning. These techniques can double or triple usable context without larger models by making better use of available space in long-running agents and large document processing.

Enable compaction when context usage approaches limits during extended agent runs. Summarize accumulated context into concise summaries and refresh the context window, improving throughput and reducing latency in ongoing conversations and multi-step tasks.

Observation masking replaces verbose outputs with compact references to reduce token consumption. KV-cache optimization reuses cached computations across repeated queries, avoiding redundant processing and lowering cost in long-running agent systems.

Yes. Context partitioning splits work across isolated contexts for scalability, enabling agents to process large documents by breaking them into manageable chunks while maintaining budget tracking and trigger-based optimization across partitions.

Budget management allocates and monitors context quotas, triggering optimizations automatically when usage approaches limits. This prevents overflow in long-running systems by enforcing explicit guardrails and enabling proactive compaction or partitioning.