LangChain Introduces Fork and Isolated Context Modes for Multi-Agent Systems

LangChain published an article on September 8 outlining a new context management feature called 'context modes' in its deepagents library, designed to improve how subagents receive information from supervisor agents. The library now offers two modes: 'isolated,' where a subagent starts with a blank context and sees only its assigned task, and 'fork,' where the subagent inherits the supervisor's full conversation history before continuing its work. The fork mode is recommended for worker subagents executing tasks the supervisor has already researched, as it avoids redundant information gathering and enables better prompt caching to reduce token costs. Conversely, the isolated mode suits verifier or researcher subagents that need an unbiased perspective, since exposure to the supervisor's prior reasoning can skew their independent judgment. The framework reframes context management as an organizational decision, comparing it to a manager choosing whether to hand an employee a full notebook of notes or only a brief summary before assigning work.