LangGraph & State Machines

Why LangGraph?

Traditional agent executors are "while loops" that run until a stopping condition. They are hard to debug, hard to control, and prone to infinite loops. LangGraph models agent workflows as state machines (directed graphs).

Core Concepts

• State: A Python dict or Pydantic model holding the current context. Passed between nodes.
• Nodes: Functions that receive state, perform actions (LLM call, API call), and return updated state.
• Edges: Rules dictating which node runs next.
• Conditional Edges: Dynamic routing based on state — enables loops and branching.

Human-in-the-Loop (HITL)

LangGraph can pause execution and wait for human approval before proceeding. Combined with a durable Checkpointer (Postgres), execution state is saved before destructive actions. This unlocks native approval workflows for payments, deployments, and refunds.

DSPy: Compile Prompts, Don't Hack Them

Senior engineers use frameworks like DSPy to treat prompts as learnable parameters. DSPy separates logical execution from instructions, dynamically compiling and optimizing the highest-accuracy few-shot combinations against test schemas.

1from langgraph.graph import StateGraph, END
2from typing import TypedDict
3
4class AgentState(TypedDict):
5    messages: list[str]
6    needs_human_review: bool
7
8def llm_node(state: AgentState):
9    return {"messages": ["LLM: I propose deleting the database."]}
10
11def tool_node(state: AgentState):
12    return {"messages": ["Tool: Action Executed."]}
13
14def should_continue(state: AgentState):
15    if state.get("needs_human_review"):
16        return "human_review"
17    return "tools"
18
19workflow = StateGraph(AgentState)
20workflow.add_node("llm", llm_node)
21workflow.add_node("tools", tool_node)
22
23workflow.set_entry_point("llm")
24workflow.add_conditional_edges("llm", should_continue, {
25    "human_review": END,
26    "tools": "tools"
27})
28workflow.add_edge("tools", END)
29
30app = workflow.compile()
31result = app.invoke({"messages": [], "needs_human_review": False})