An agent is a graph

Short summary

A tutorial on building a Claude-based chatbot called Claudius, treating the agent as a LangGraph state machine with three nodes: context loading, model invocation, and tool execution. The key design decision is passing the model per-run via configurable channels rather than compiling it into the graph, enabling seamless mid-conversation model switching (e.g., Haiku to Sonnet) without losing state. Includes TypeScript code for the graph construction and agent node implementation.

• •Agent chat engine is built as a LangGraph state machine with load_context, agent, and tools nodes
• •Model is chosen per-run via configurable channels, not compiled into the graph
• •Enables mid-conversation model switching (e.g., Haiku to Sonnet) without losing conversation state

Generated with AI, which can make mistakes.