context-optimization

Reduce token usage in long-running AI agent sessions by applying context optimization techniques like KV-cache optimization, observation masking, compaction, and context partitioning to maintain signal with fewer tokens.

Context partitioning splits work across sub-agents when a single window cannot hold the full problem, coordinating results to produce a final answer. It is needed for multi-turn conversations exceeding context limits.

Maximize prefix stability to reuse cached tokens by structuring prompts for KV-cache optimization, ensuring repeated requests reuse cached tokens to reduce cost and latency in AI agent sessions.

Observation masking works by replacing verbose tool outputs with compact references while keeping the ability to retrieve full content later, effectively reducing token usage in document-heavy tasks.

The best way to summarize older context when approaching token limits is compaction, which summarizes older context while preserving decisions and critical state to extend effective context capacity.

You can configure budgets and thresholds for context compaction and partitioning to manage strict budget constraints in long-running agent systems handling large documents or multi-turn conversations.