ADR-001: LangGraph for Agent Orchestration
Status
Accepted
Context
RavenmaskOS requires an AI agent framework that can:
- Handle complex multi-step reasoning
- Support tool use and function calling
- Maintain conversation state and memory
- Enable modular, composable agent workflows
We evaluated several frameworks: LangChain agents, AutoGen, CrewAI, and LangGraph.
Decision
We chose LangGraph as the orchestration framework for the Norns agent.
LangGraph provides:
- Graph-based workflow definition
- Built-in state management
- Native tool/function calling support
- Streaming responses
- Human-in-the-loop capabilities
- Integration with LangChain ecosystem
Consequences
Positive
- Clean separation of agent logic into graph nodes
- State persistence across conversation turns
- Easy to add new tools and capabilities
- Strong debugging with LangSmith/Langfuse integration
- Active development and community support
Negative
- Learning curve for graph-based thinking
- Some complexity for simple use cases
- Dependency on LangChain ecosystem
Alternatives Considered
LangChain Agents (legacy)
Simpler but less flexible. Deprecated in favor of LangGraph.
AutoGen
Good for multi-agent scenarios but more complex setup. Less suitable for single-agent tool use.
CrewAI
Role-based agent framework. Better for collaborative agent scenarios than personal assistant use case.