Autonomous context compression

A new tool has been added to LangChain's Deep Agents SDK that allows AI models to autonomously compress their own context windows. This feature addresses the limitations of fixed-threshold compaction, which can interrupt complex tasks, by letting the agent decide the most opportune time to summarize its working memory. Opportune moments for compaction include finishing a task, starting a new one, or after extracting key information from a large amount of text. This approach reflects a design philosophy of giving models more control over their own processes rather than relying on rigid, hand-tuned rules in the agent harness.