When does a LangGraph conditional-edge router beat LangChain `RouterRunnable`?

Short answer·Hard·4.0 · 0·~3 min·Asked atMicrosoft Cursor Perplexity

Attempt it

Both `RouterRunnable` (LangChain LCEL) and a conditional-edge router (LangGraph) let you pick which downstream chain runs based on the input. When should you reach for the LangGraph version instead of the LCEL one?

Free · 2 AI evals / day

Both tools answer the same surface question. Given an input, decide which downstream component runs. The difference is the shape of the workflow they live inside. `RouterRunnable` is part of LangChain's LCEL. LCEL composes Runnables into a DAG. Every step runs once, in order, and the graph has no cycles. Inside that DAG, RouterRunnable is the place you classify the input and pick one of several chains to forward to. It works perfectly for one-shot routing: classify the question, pick a chain, return the chain's output, done. LangGraph's conditional-edge router lives in a StateGraph, which is allowed to have cycles. After any node runs, a conditional edge can route to one of several next nodes. Including back to a node that already ran. Combined with a typed state that all nodes share, and a checkpointer that persists state at every step, you get capabilities LCEL cannot match: loops, shared state across branches, pause and resume runs, and human in the loop interrupts. The rule of thumb: one decision in a feed-forward pipeline is RouterRunnable. A decision that repeats in a loop, or shares state across branches, is LangGraph.

**RouterRunnable's sweet spot is the one-shot decision inside an LCEL pipeline.** Classify the question, pick a chain, return its output. The pipeline is a DAG, the decision runs once, and there is no need to come back after the branch ran. It is the right tool when there is no loop, no shared mutable state across branches, and no need for resumability. **Reach for LangGraph when at least one of these is true:** - **Routing repeats in the same run.** After every tool call, the agent decides what to do next. Call another tool, synthesize, or escalate. LCEL is a DAG; it cannot express cycles. LangGraph's conditional edges can route back into the same node. - **Branches share state.** A typed `State` (TypedDict) with reducers gives every node access to the same evolving data. Chat messages, tool results, intermediate plans, scratchpad. LCEL passes outputs forward; sharing state across branches requires plumbing you write yourself. - **Checkpointing matters.** LangGraph checkpoints state at every super-step. You can pause the run, inspect state, edit it, resume from where you stopped, or replay from a prior checkpoint for debugging. RouterRunnable inside a Runnable chain has no native checkpoint. - **Parallel fan-out with merge.** LangGraph supports super-step parallelism: multiple branches execute in the same step and their results are merged via reducers. LCEL has `RunnableParallel` for static fan-out but no clean primitive for fan out then route then merge with state. **Heuristic.** One decision in a pipeline → RouterRunnable. A decision in a loop, or with shared evolving state, or that needs to pause for a human → LangGraph.

Quick reference

TL;DR

RouterRunnable is for one-shot routing inside a DAG; LangGraph wins when routing repeats in a loop, branches share state, or runs need checkpointing.

Memory aid

Easy to grasp

Picture a phone tree at a clinic. RouterRunnable is the menu you hear once when you call. 'press 1 for billing, 2 for appointments, 3 for records'. You press, get sent to one department, they handle it, the call ends. One decision, one path. LangGraph routing is the back office where staff hand a chart between desks several times in the same visit. The triage nurse sends you to radiology, who sends you back to the doctor with the X-ray, who sends you to the pharmacy. Many decisions, shared chart, sometimes the doctor pauses for a senior consult before continuing. Use the phone tree for single-decision flows; use the back-office model when the chart keeps moving.

Key concepts

LLM Frameworks Agents

Detailed answer & concept explanation

~5 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Quick scan · 2–4 min

Everything important, quickly.

5 min: DAG vs graph topology, where RouterRunnable wins, the four LangGraph capabilities, and the heuristic for picking between them.

Side-by-side

Capability	RouterRunnable (LCEL)	LangGraph conditional edges
Workflow shape	DAG, no cycles	Graph with cycles
Routing decisions per run	One	Many (after each node)
Shared state across branches	Pass outputs forward	Typed State with reducers
Checkpointing	Not native	Built into runtime
HITL interrupts	Not native	interrupt_before / after any node
Parallel fan-out + merge	RunnableParallel (static)	Super-step + reducer (dynamic)
Best fit	One-shot classify and dispatch	Agent loops, multi-step routing, HITL

Real-world examples

Real products, models, and research that use this idea.

LangChain's own templates use RouterRunnable for 'pick the right RAG chain' flows and LangGraph for 'tool-using agent' flows.
Replit's coding agent uses LangGraph because the routing decision repeats after every tool call and needs checkpointing for HITL approval.
Klarna's customer-support classifier uses LCEL RouterRunnable to dispatch to one of several response chains in a feed-forward pipeline.
LangGraph Studio visualises conditional edges as routing arrows with the predicate inline; the same flow in pure LCEL is invisible inside a RunnableLambda.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QCould you implement a LangGraph-style loop using only LCEL?

QHow does LangGraph's State + reducers compare to a Redux store?

QWhen would you choose RunnableBranch over RouterRunnable?

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Using RouterRunnable for an agent loop and discovering halfway through that LCEL cannot cycle. Forcing a rewrite into LangGraph after the codebase is committed.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

References

Primary sources. Skim if you want the original framing.