Explain the structural difference between a LangGraph StateGraph and a standard DAG based LLM pipeline. What specific graph feature enables the agent's 'retry until done' pattern?
A DAG forbids cycles, so it cannot loop back; StateGraph allows directed cycles, and a conditional edge routes on state to repeat the reason step or terminate.
Think of a board game with a track of squares. A normal pipeline is a track where you only move forward, one square at a time, until you reach the end. You can never go back. That is fine for a recipe you follow once, but an agent does not know in advance how many tries it needs. LangGraph lets you draw an arrow from a later square back to an earlier one, so the player can loop around again. At the end of each loop there is a decision square. Look at the current situation, and if the job is done, walk off the board, otherwise take the arrow back and try again. The decision is written on the board itself, not hidden in someone's head, so anyone watching can see the whole game and where the player keeps getting stuck.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
A StateGraph is LangGraph's answer to a structural limitation in classic LLM pipelines. Most chain frameworks compile your workflow into a directed acyclic graph. Acyclic is the operative word: every edge points forward, the runtime executes nodes in topological order, and no node can run twice. That model fits a recipe like retrieve, then generate, then format the output. It cannot fit an agent.
An agent is fundamentally a loop. It reasons about the next action, calls a tool, observes the result, and frequently has to reason again with that new observation. The number of turns is not known until runtime, because it depends on what the tools return. Encoding that requires an edge that points backward, from a later node to an earlier one, which is exactly the cycle a DAG forbids. StateGraph removes that restriction and makes the loop a first-class, inspectable object rather than control flow hidden inside application code.
Why a DAG structurally cannot loop
A directed acyclic graph guarantees that following edges never returns you to a node you have already visited. This is not a stylistic preference; it is what makes topological ordering possible. The runtime can lay all nodes out in a line such that every edge points forward, then execute them left to right, once each. Latency and cost are bounded by the number of nodes.
Add a single edge from a later node back to an earlier one and the graph becomes cyclic. There is no longer any topological order, because the back edge creates a node that depends, transitively, on itself. The execution model breaks. This is why a chain DSL that compiles to a DAG can express retrieve then generate, but cannot express reason, act, observe, reason again.
The agent loop needs precisely that forbidden back edge. After executing a tool, control must be able to return to the reasoning step so the model can incorporate the observation. No acyclic structure can represent this, regardless of how the nodes are arranged.
It is worth being precise about why the common workaround falls short. You can simulate a loop by wrapping a DAG in an external while loop in application code, calling the compiled chain over and over until some Python condition flips. That technically runs the agent, but the loop itself is now invisible to the framework. The graph the framework knows about is still acyclic, so the cycle, its termination condition, and its per-turn state all live outside the graph where no tooling can see them. The structural point is that the iteration must be part of the graph, not bolted on around it.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- LangGraph's create_react_agent compiles a reason node and a tool node joined by a conditional edge, producing the canonical observe reason act loop as an inspectable state machine.
- LangGraph checkpointers like MemorySaver, SqliteSaver, and Postgres snapshot state after each node, powering pause and resume and human approval flows in agents built on Claude Opus 4.7 or GPT-5.5.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does a reducer change the semantics of a state update when two nodes write the same key concurrently?
Explain that without a reducer the last write wins and clobbers prior values, while a reducer like add_messages defines a merge, so concurrent branches append rather than overwrite, making fan-out and fan-in deterministic.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Saying StateGraph is just a nicer wrapper around a while loop. The real difference is that the loop becomes a typed graph object you can inspect, checkpoint, resume, and interrupt.
60 second bullets to scan on the way to the call.
State why topological execution is undefined on a cyclic graph.
Explain how a back edge turns a DAG into a cyclic graph.
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.