LangChain has introduced two context modes for subagents in its Deep Agents framework. The existing isolated mode starts a delegated task with a fresh context window and only the supervisor’s task description. A new fork mode passes along the supervisor’s conversation state before the subagent begins.
Forking is intended for workers that continue an investigation already performed by the supervisor. Those workers can reuse gathered evidence rather than reading the same files or reconstructing earlier decisions, and shared prefixes can benefit from prompt caching. When the subagent finishes, only its final response returns to the supervisor as the tool result.
Isolation remains the better fit when independence matters. A reviewer checking a code change may be biased by inheriting the original diagnosis, while parallel researchers on self-contained questions do not need a large shared history. LangChain’s distinction gives developers control over that trade-off instead of using one context policy for every delegated task. Forking can reduce duplicate work, but it also sends more existing context into each subagent.