●builderYou can use model-driven compaction to maintain long-running agentic loops without losing task-relevant information.
●researcherThis method optimizes context management by tying compaction triggers to trajectory structure rather than fixed token counts.