Your team is running an agent in production on `AgentExecutor`. At what point should you graduate to a `LangGraph StateGraph`, and what specifically does the StateGraph give you that the executor cannot?
Free · 2 AI evals / day
**LangChain's `AgentExecutor`** runs a tool-using agent as a hidden `while` loop: ask the LLM what to do, parse the response, call the tool, append the result, repeat. Great for prototypes. The trouble starts in production.
Three production needs `AgentExecutor` cannot serve natively. First, **debuggability**: when a loop branches in a weird way, you can read traces but you cannot easily rewind to step 5 and try again with a tweaked input. Second, **durable state**: if the worker crashes mid-loop, the run is lost. Third, **human in the loop**: there is no clean way to pause the loop before an irreversible action, get approval, and resume.
**LangGraph's `StateGraph`** addresses all three. Nodes are your steps. Edges (and conditional edges) are your routing. A checkpointer persists state at every step, so crashes resume from the last checkpoint. `interrupt_before(node)` halts the graph and lets a human approve before the action fires. Time-travel debugging works because every step is checkpointed.
Graduate when you hit one of those three triggers, not before. For a 50-line prototype, the executor is still fine.
**Graduate when the hidden while loop becomes an obstacle, not before.** Three concrete production triggers force the move from AgentExecutor to StateGraph.
**One: debuggability.** A flaky loop branches in unexplained ways. With AgentExecutor the loop body is internal, you can log inputs and outputs but you cannot rewind to step 5 with a mutated state and replay. StateGraph checkpoints every step, so `graph.get_state_history()` plus `graph.update_state()` lets you fork and re-run from any prior state.
**Two: durable state.** A worker crashes mid-loop. AgentExecutor loses the run. StateGraph with a Postgres saver (or a Redis saver) resumes from the last checkpoint when you re-invoke with the same thread id. The graph picks up exactly where it was.
**Three: human in the loop.** An action is irreversible (send email, place trade, run DB migration) and needs human approval. AgentExecutor has no primitive for this. StateGraph compiles with `interrupt_before=["send_email"]`; the graph halts at that node, your app surfaces the pending state to a human, on approval you resume with `graph.invoke(None, config)`.
**Below those triggers, stay on AgentExecutor or `create_react_agent`.** The graph machinery is verbose, you write nodes, edges, conditional edges, a typed state with reducers, and wire it all up. For a 50-line prototype that calls one tool, the executor is still right. The graduation is a deliberate response to a real production need, not a cosmetic upgrade.
Quick reference
TL;DR
Graduate when you need debuggability, durable state, or human in the loop approval, StateGraph makes the hidden AgentExecutor while-loop into an inspectable, checkpointable, interruptible graph.
Imagine driving with a closed-eyes navigation app vs an open map. AgentExecutor is the closed app, it gets you there, but if it takes a weird detour you cannot see which turn it picked or why, and if your phone dies mid-route you start over. StateGraph is the open map: every turn is a visible step, you can pause the trip, hand the wheel to a passenger for one decision, and even rewind to a previous junction and try a different road. For a quick errand the closed app is fine. For a complex delivery route with handoffs and recovery requirements, you want the map.
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.
A 50-line prototype agent with one tool and no need for recovery or approval steps. Graph machinery is overhead. AgentExecutor or the `create_react_agent` helper is shorter, simpler, and just as fast for the prototype phase.
A stateless single-turn classification or extraction call. There is no loop to make explicit. A plain LCEL chain (prompt then model then parser) is the right abstraction.
A team that has standardized on a different orchestration framework (AutoGen, CrewAI, OpenAI Agents SDK) and only has one or two agents. Mixing two orchestration frameworks fragments observability and adds mental overhead. Commit to one unless the workload genuinely demands two.
Quick scan · 2–4 min
Everything important, quickly.
State the trigger framework (crash recovery, HITL, debuggable replay), explain what each maps to in StateGraph (checkpointer, interrupt_before, state history), call out the verbosity tradeoff, name when the executor is still right, and close with the concrete migration-decision heuristic.
Real-world examples
Real products, models, and research that use this idea.
1Replit's Agent migrated from a custom executor to LangGraph because IDE-disconnect and reconnect required durable session state across restarts.
2Norway's tax authority pilot used LangGraph `interrupt_before` to gate citizen-facing actions on a human caseworker's approval.
3LangChain's own production guidance explicitly recommends LangGraph as the migration target for `AgentExecutor`-based agents going to production.
4Anthropic's Claude Code uses graph-based agent orchestration to gate file-write and shell-execution actions on user approval, the same pattern LangGraph supports natively.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWalk through how a `PostgresSaver` checkpointer enables crash recovery in a real deployment.
Sign in or go Pro to see the approach for this follow-up.
QWhat's the practical difference between `interrupt_before` and `interrupt_after`, and when do you reach for each?
Sign in or go Pro to see the approach for this follow-up.
QYour `AgentExecutor`-based agent works fine on 50 QPS and the team wants to know if migrating is premature optimization. How do you frame the decision?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Graduating to LangGraph for a 50-line agent that calls one tool. The verbosity is overhead unless you have a concrete trigger like crash recovery or HITL.