A better handoff: deciding what a subagent should remember.
LangChain describes isolated and forked context modes in Deep Agents. Workers can inherit earlier investigation, while reviewers can begin independently. The choice changes how teams manage repeated work and bias.
LangChain's September 8 engineering post explains two context modes for subagents in Deep Agents. Isolated subagents receive a focused assignment in a fresh context. Forked subagents inherit the supervisor's conversation, which can avoid repeating investigation the supervisor has already completed.
The distinction is useful when designing agent teams. An implementation worker may benefit from the history of a diagnosed issue. A reviewer may instead need the evidence and acceptance criteria without the preceding reasoning. LangChain notes that the right choice depends on the subagent's relationship to the work.
Our view: treat context sharing as an explicit design decision. More history is not automatically better. Test whether inherited context improves completion time without causing workers to repeat an earlier mistake. Evaluate reviewers on their ability to find real defects, and make sure necessary evidence is available even when the prior conversation is withheld.
give workers useful history and reviewers room to disagree.
Read the primary source. Editorial interpretation is our own.