LangChain Introduces 'forked subagents' to deepagents to Eliminate Context Duplication and Improve Performance

LangChain has introduced the forked subagents feature in the latest version of deepagents. Previously, subagents started with a fresh context window, leading to inefficiencies such as redundantly performing context-gathering tasks like file reading that the supervisor had already completed.

How Forked Subagents Work

The new fork mode propagates the supervisor's entire conversation state to the subagent. The subagent operates as a continuation branched from the supervisor's thread, with the supervisor's instructions appended. This enables the use of prompt caching, making it faster and cheaper than isolated subagents.

Context Mode Usage Strategy

Depending on the subagent's role, you can selectively use isolated and fork modes.

• Worker Agent (fork): Used when continuing a task after the supervisor has already gathered context. It is efficient because it can resume from the point where investigation was interrupted.
• Verifier Agent (isolated): Suitable for verification tasks requiring independent judgment. Inheriting the supervisor's reasoning can cause an anchor effect, so only relevant materials are passed.
• Researcher Agent (isolated): Used for investigating independent questions. It prevents duplication of supervisor history during parallel execution.

deepagents is a framework developed based on the LangChain team's experience with thousands of team tasks, supporting both Python and TypeScript.