LangChain 1.3.3 adds Human-in-the-Loop middleware and subagent improvements
New version enhances human oversight and subagent coordination for LLM workflows.
LangChain 1.3.3, released on June 2, 2025, brings two notable features to the popular LLM application framework. First, the new HumanInTheLoopMiddleware allows developers to define precisely when and how human approval is required during agent runs. It introduces an interrupt_mode parameter and a when predicate, enabling conditional halting of execution based on custom logic—such as high-cost API calls or sensitive actions. This gives teams building production-grade agents a structured way to integrate human oversight without breaking the flow.
Second, the release enhances subagent management with the ability to project subagent runs onto a typed run.subagents channel. This means outputs from child agents can be more easily captured, typed, and routed within the parent execution graph. Additionally, the langgraph dependency is bumped to version 1.2.4 and its version constraint is loosened for greater compatibility. These changes are incremental but meaningful for developers orchestrating multi-agent systems where reliability and observability are critical.
- HumanInTheLoopMiddleware adds interrupt_mode and when predicate for conditional human approval
- Project subagent runs onto typed run.subagents channel for better subagent output handling
- Langgraph dependency bumped to 1.2.4 with loosened version range for easier updates
Why It Matters
Enables safer, more controlled LLM agent deployments with human oversight and structured subagent communication.