Agentsop Langgraph

LangGraph · SOP

Source posture: every non-trivial claim is cited inline. Citations use short tags like [lc-docs], [lc-blog/interrupt], [gh/6731], [zenml/uber] — resolve them against references/*.md for the full URL.

何时激活 (Activation Rules)

Activate this skill when any of the following triggers fire:

• The task mentions LangGraph, StateGraph, MessageGraph, create_react_agent, interrupt(, Command(resume=, add_messages, checkpointer, PostgresSaver, Send(, or entrypoint / task decorators.
• The user wants to build a stateful agent (memory across turns, long-running, must survive a process crash) — LangGraph's stated sweet spot [lc-docs/why-langgraph].
• The user wants human-in-the-loop (approve a tool call, edit state, multi-turn validation) — LangGraph offers a first-class interrupt() primitive that competitors require "duct-taping" to achieve [bswen/hitl].
• The user is hitting GRAPH_RECURSION_LIMIT errors, infinite loops, or InvalidUpdateError on parallel branches — these are LangGraph-specific failure modes with known fixes [lc-docs/errors] [cheatsheet/gotchas].
• The user is choosing between LangGraph and CrewAI / AutoGen / OpenAI Swarm / raw LangChain — section 生态对照 gives the decision matrix.
• The user is migrating an existing LangChain chain or a hand-rolled while-loop agent to something durable and observable.

Do not activate if the task is a single LLM call, a one-shot RAG query, or a stateless tool pipeline — Sec. 反模式 explains why graphs are overkill there.

核心心智模型 (Core Mental Model)

LangGraph is a state machine, not a chain. The cleanest one-liner from the 2026 docs: "If chains were about passing outputs between steps, graphs are about maintaining and evolving a shared state over time" [eastondev/2026]. Pre-LLM analog: think BPMN / finite state machine / Pregel-style "supersteps", not a Unix pipe. The official position is even more reductive: LangGraph is "a deterministic execution engine for AI reasoning workflows" [eastondev/2026].

Three load-bearing concepts ride this model:

1. State is the single source of truth. All nodes read from and write to one shared, typed object (TypedDict / Pydantic / dataclass). A node returns a partial update, never a mutation. How updates merge into state is governed by reducers, declared via Annotated[list[Msg], add_messages] etc. Missing a reducer on a key that two parallel nodes both write to triggers InvalidUpdateError — reducers are mandatory for parallel writes [cheatsheet/gotchas]. The reducer system is what lets the graph be composable, replayable, and crash-safe.

2. Checkpoints make state durable. After every superstep, the full state is snapshotted into a checkpointer (SQLite for local, Postgres for production, Redis for fast TTL'd swarms) [lc-docs/persistence] [redis/checkpoint]. This single property is what unlocks the headline features: durable execution that "persists through failures and resumes from their exact stopping point", time-travel debugging (replay or fork from any checkpoint), and human-in-the-loop (a thread can sit interrupted for hours and resume cleanly) [gh/langgraph-readme] [dragonforest/timetravel].

3. Graph topology is just routing logic over state. Edges are static (always go to N), conditional (a function reads state and picks a next node), or dynamic via the Send API (a routing function returns a list of Send objects to spawn variable-count parallel workers) [deepwiki/mapreduce]. This is where LangGraph diverges from CrewAI's role-based crew and AutoGen's conversational pattern — control flow is explicit, not emergent from chat history.

The OS-level claim: "2026 is the year of Stateful Orchestration" [eastondev/2026]. LangGraph bet that production agents need persistence, explicit control flow, and observability more than they need elegance. That bet is paying off (Klarna serves 85M users on it, Replit pushed it so hard LangSmith had to be rewritten to ingest the traces) — but the cost is verbosity that frustrates anyone trying it on a toy problem [lc-blog/production] [duplocloud/compare].

SOP 工作流 (Agentic Protocol)

A coder agent should walk this protocol top-down. Each step has a decision gate — if the answer is "no" or "not yet", stop and reconsider before adding graph complexity.

Step 1 · Decide whether a graph is actually warranted

Gate questions:

• Does the workflow have ≥1 cycle (tool-call → reflect → retry)?
• Does it need to survive a crash mid-execution?
• Will a human need to inspect or override state mid-run?
• Are there ≥2 specialized agents that hand off?

If all four are no, use a plain RunnableSequence or raw API calls and exit. Over-graphing simple flows is the #1 anti-pattern [swarnendu/best].

Step 2 · Pick the API surface

Default to create_react_agent and graduate to StateGraph only when you need parallel nodes, supervisor-worker patterns, custom retry logic, or complex branching [agentsindex/v1].

Step 3 · Design the state schema before writing nodes

The state schema is "the most critical design component" [bharatraj/state]. Discipline:

• Use TypedDict for ergonomics, Pydantic only when validation matters.
• Every key that may be written in parallel gets an explicit reducer (add_messages, operator.add, or custom) — otherwise plan for it to be overwritten last-write-wins.
• Keep state lightweight and serializable — it gets pickled to the checkpointer on every superstep [bharatraj/state].
• Treat each node like a pure function: return a partial state update, do not mutate inputs [swarnendu/best].

Step 4 · Choose the multi-agent topology

Decision tree, sourced from LangChain's own benchmark [lc-blog/benchmark]:

Is there exactly one "user-facing" persona?
├─ YES  → Supervisor pattern (single supervisor, sub-agents are tools)
│        - Highest token cost (supervisor "translates" sub-agent output)
│        - Safest with third-party agents
│        - LangChain's *current recommended default*
└─ NO   → Do sub-agents know about each other?
         ├─ YES → Swarm pattern (dynamic handoff, last-active agent remembered)
         │       - Lower tokens than supervisor (no translation step)
         │       - Slightly higher accuracy in the τ-bench retest
         │       - Bad fit for third-party agents
         └─ NO  → Hierarchical Teams (supervisor-of-supervisors)
                 - Use only when ≥6 specialists need grouping

Concrete bench finding: swarm "slightly outperformed supervisor across all scenarios"; supervisor "consistently uses more tokens than swarm" because of the telephone-game translation overhead [lc-blog/benchmark]. LangChain's own response was to fix the supervisor (remove handoff messages, add a forwarding-messages tool, tune tool names) for "a nearly 50% increase in performance" [lc-blog/benchmark].

Step 5 · Add human-in-the-loop only on irreversible actions

Use interrupt(value) at the node that would perform the high-blast-radius operation; resume with Command(resume=...) [lc-blog/interrupt]. Four canonical patterns [lc-blog/interrupt]:

1. Approve / Reject — review a critical step before it runs.
2. Review & Edit State — human corrects or augments mid-run.
3. Review Tool Calls — oversee LLM-requested actions.
4. Multi-turn Conversation — back-and-forth in a multi-agent setup.

Rule of thumb: "interrupt on irreversible, high-blast-radius actions only — not on every step" [bswen/hitl]. Side effects (DB writes, API calls) must go after the interrupt or in a downstream node — placing them before causes unwanted re-execution on resume [cheatsheet/gotchas].

Step 6 · Pick the checkpointer to match the durability requirement

Run checkpointer.setup() as a CI/CD migration, never inside app runtime [bswen/hitl]. Implement a TTL sweep for interrupted-but-never-resumed threads (e.g., abandon after 24 h) — otherwise state accumulates indefinitely [bswen/hitl].

Step 7 · Add observability + bounded loops before shipping

• Wire LangSmith from day one — replaying a checkpoint locally only goes so far; production needs the trace UI [swarnendu/best].
• Set a deliberate recursion_limit (default 25); raise it via graph.invoke({...}, {"recursion_limit": 100}) only after confirming the loop can terminate [lc-docs/errors].
• Treat recursion_limit as a safety net, not control flow. Hitting it means the conditional edge logic is wrong, not that the limit is too low [cheatsheet/gotchas].

操作模型 (Operation Models)

Each operation is a primitive a coder agent can invoke. Format: Trigger → Action → Output → Evidence.

OP-1 · Bootstrap a ReAct agent in <10 lines

• Trigger: User says "make me an agent that uses tool X" with no other requirements.
• Action: Call from langgraph.prebuilt import create_react_agent; pass model + tools list. Skip StateGraph entirely.
• Output: A compiled graph supporting .invoke() / .stream() with built-in message history.
• Evidence: [agentsindex/v1] "Start with create_react_agent for any standard tool-calling agent."

OP-2 · Promote a prebuilt agent to a custom StateGraph

• Trigger: The prebuilt agent needs parallel branches, a supervisor, custom retry, or a non-message state field.
• Action: Re-implement with StateGraph(MyTypedDict), manually add the LLM node, tool node, and conditional edge that routes on tool_calls.
• Output: A graph with explicit topology and full control.
• Evidence: [agentsindex/v1] "If you find yourself needing parallel node execution, a supervisor-worker pattern, custom retry logic, or complex branching, migrate to a manual StateGraph."

OP-3 · Add a reducer to fix InvalidUpdateError

• Trigger: Two nodes write the same state key in parallel and the graph raises InvalidUpdateError.
• Action: Replace key: list[X] with key: Annotated[list[X], operator.add] (or add_messages for chat).
• Output: Parallel writes merge instead of conflicting.
• Evidence: [cheatsheet/gotchas] "Reducers are mandatory, not optional, for parallel execution."

OP-4 · Spawn dynamic parallel workers with Send

• Trigger: At runtime, the agent needs to fan out a variable number of parallel tasks (e.g., one summarizer per retrieved doc).
• Action: From a conditional edge, return [Send("worker", {"chunk": c}) for c in state["chunks"]]. The worker uses its own state schema, and results are reduced back via operator.add.
• Output: True dynamic map-reduce, decided per-invocation.
• Evidence: [deepwiki/mapreduce] "Send allows a conditional edge function to schedule a node with a custom state…the primary mechanism for dynamic fan-out."

OP-5 · Insert a human approval gate

• Trigger: A node would perform an irreversible action (charge a card, send email, run SQL DELETE).
• Action: Inside the node, call decision = interrupt({"proposed": payload}) before the side effect. Compile the graph with a checkpointer. The caller resumes with graph.invoke(Command(resume="approve"), config).
• Output: Thread pauses, state persists, resumes with the human's decision threaded back into the node.
• Evidence: [lc-blog/interrupt] four-pattern table + [bswen/hitl] side- effect ordering.

OP-6 · Encapsulate a multi-step subflow as a subgraph

• Trigger: A logical chunk (a research agent inside a larger workflow, a reusable validation pipeline) deserves its own state and lifecycle.
• Action: Compile the child graph, add it as a node: parent.add_node("research", research_subgraph). If schemas differ, wrap the call in a node function that translates between schemas.
• Output: Reusable, independently testable module; child state isolated unless keys overlap.
• Evidence: [lc-docs/subgraphs] [deepwiki/subgraphs] "subgraph state is only accessible when the subgraph is interrupted" — accept the debug cost.

OP-7 · Switch to durable Postgres checkpointing

• Trigger: Moving from notebook to production OR multi-user OR thread state must survive deploys.
• Action: from langgraph.checkpoint.postgres.aio import AsyncPostgresSaver, pass to .compile(checkpointer=...). Run await saver.setup() in a migration job, not at app boot.
• Output: Threads survive crashes, multi-replica reads work, ACID guarantees for state transitions.
• Evidence: [lc-docs/persistence] PostgresSaver "ideal for using in production"; [bswen/hitl] "handle this as part of a CI/CD migration script…not inside the primary application runtime."

OP-8 · Stream tokens to the UI without sacrificing graph observability

• Trigger: Need character-by-character UX and server-side debugging.
• Action: Use graph.stream(input, stream_mode=["messages", "updates"]) — messages yields LLM tokens, updates yields state diffs. For intra-tool progress, emit via stream_mode="custom".
• Output: User sees streaming tokens; server logs structured state diffs.
• Evidence: [lc-docs/streaming] five modes — values, updates, messages, custom, debug.

OP-9 · Bound an agent loop without papering over with recursion_limit

• Trigger: Agent hits GRAPH_RECURSION_LIMIT (e.g., text-to-SQL retrying the same broken query).
• Action: Read the conditional-edge logic — the real fix is an exit condition counting retries in state and routing to END after N attempts. Bump recursion_limit only as a temporary diagnostic.
• Output: Graceful termination on persistent failure.
• Evidence: [lc-docs/errors] "Check your logic for infinite loops"; [cheatsheet/gotchas] "Hitting the limit indicates an underlying design flaw"; concrete bug case [gh/6731].

OP-10 · Time-travel debug a failed production run

• Trigger: A production thread produced a wrong answer; need to see what state the LLM saw at step 7 and try a different prompt.
• Action: Fetch the thread's checkpoint history via graph.get_state_history(config), pick the checkpoint, invoke with config={"configurable": {"thread_id": ..., "checkpoint_id": ...}}. Modify state with graph.update_state(...) to fork.
• Output: Reproduce + fork a past run without re-paying for LLM calls on replayed nodes.
• Evidence: [dragonforest/timetravel] "replay…the agent knows that this checkpoint has already been executed and will just display the historical output instead of making new LLM calls."

困境决策案例 (Dilemma Cases)

Case 1 · "Text-to-SQL agent loops forever until recursion limit"

• 困境: A team built a text-to-SQL agent on LangGraph 1.0.6. When the Databricks query returned an error, the agent retried the same broken query 20 times until GRAPH_RECURSION_LIMIT fired. It had worked on 0.6.x [gh/6731].
• 约束:
  • Cannot pin to old version (security fixes in 1.x).
  • Maintainer marked the issue "not planned" — no official patch coming [gh/6731].
  • Business needs the agent live.
• 决策步骤:
  1. Reject the temptation to just raise recursion_limit — that masks the bug and burns Databricks quota [cheatsheet/gotchas].
  2. Add an explicit retry counter to state: retries: Annotated[int, operator.add].
  3. After the tool node, increment counter on SQL error; in the conditional edge, route to END (or a "give up and ask user" node) once retries >= 3.
  4. Surface the failure mode by storing the last error message in state and letting the LLM see it on the next loop — the content of the error usually informs whether to retry or abandon.
  5. Add a regression test that injects a permanent SQL error and asserts the graph terminates within 3 iterations.
• 结果: Bounded retries, observable failure, no quota blow-out. The community-acknowledged fact is that LangGraph's recursion limit is a safety net, not control flow — exit conditions are the real fix [lc-docs/errors].
• 可提取的操作: OP-9. Always bake a retry/exit counter into state for any cyclic graph. Trust nothing the LLM does to terminate itself.

Case 2 · "Supervisor vs. swarm in a multi-domain customer support agent"

• 困境: A team needed a customer-service agent covering retail, billing, and shipping. They started with the supervisor pattern (one router agent → three specialists). User-perceived latency was high and token cost was double what they budgeted. They wondered if swarm would be better.
• 约束:
  • The three specialists are internal — they can be aware of each other.
  • User expects continuous conversation in one domain (no constant "transferring you to...").
  • Compliance requires a single auditable agent for tool calls.
• 决策步骤:
  1. Consult LangChain's own benchmark: swarm "slightly outperformed supervisor"; supervisor uses more tokens because of the translation step [lc-blog/benchmark].
  2. Map constraints to patterns: compliance favours supervisor (single funnel); UX favours swarm (last-active agent stays active across turns).
  3. Compromise: keep supervisor topology but apply LangChain's three supervisor fixes — "removing handoff messages, forwarding messages tool, tool naming optimization" — which yielded a "nearly 50% increase in performance" in the bench [lc-blog/benchmark].
  4. Re-evaluate token cost after the fixes; if still too high and audit trail is tolerant, migrate to swarm.
• 结果: Hybrid — supervisor topology with tuned tools captures most of swarm's efficiency while preserving the single-funnel audit log.
• 可提取的操作: Don't pick supervisor vs. swarm on aesthetics — anchor on (a) whether sub-agents can know each other, (b) whether one user-facing voice is mandated. Then optimise the chosen pattern with LangChain's own published fixes before switching paradigms.

Case 3 · "Replit-scale traces overwhelmed the observability stack"

• 困境: Replit built a code-generation agent on LangGraph that "involved hundreds of steps" per run. Traces were so large that LangSmith — built by the same team — couldn't ingest or render them initially [alphabold/case].
• 约束:
  • The depth of the graph is intrinsic to the product (planning → code → tests → deploy → debug → fix).
  • Cannot trim steps without harming product quality.
  • Need debugability for a multi-agent system with HITL.
• 决策步骤 (reconstructed from the case study):
  1. Accept that pushing LangGraph to its limit means co-evolving the observability layer — Replit and LangChain iterated on LangSmith's ingestion and rendering specifically for this trace shape [alphabold/case].
  2. Use subgraphs to break "hundreds of steps" into named, navigable teams (planner-team, codegen-team, test-team) — flat graphs of that size are unreadable [deepwiki/subgraphs].
  3. Use Send API for fan-out at known parallel points (e.g., generate-then-test in parallel) so each branch is a distinct trace segment.
  4. Use HITL interrupt at the deploy boundary — humans approve a deploy plan rather than letting the agent push autonomously.
  5. Use time-travel on failed runs — fork from the last good checkpoint to test prompt variants without re-paying for upstream LLM calls [dragonforest/timetravel].
• 结果: Replit Agent ships; the friction it surfaced fed back into LangSmith improvements. The lesson: at scale, the observability tool is part of the system design, not external to it.
• 可提取的操作: For agents with > ~50 steps per run, plan the observability story alongside the graph topology. Subgraphs + Send are not optional optimisations — they are how you make the graph debuggable at production scale.

Case 4 · "Side effects before interrupt() re-executed on resume"

• 困境: A team built a payment workflow: node A charges the card, then calls interrupt() for a human to confirm the receipt. On resume, the card was charged twice because resuming a thread "re-runs the entire node function" [cheatsheet/gotchas].
• 约束: Can't disable HITL (compliance requirement); can't undo charges silently; cannot rewrite payment SDK.
• 决策步骤:
  1. Recognize the framework semantics: resume re-enters the node from the top, not from the line of the interrupt(). Treat every node body as potentially re-runnable.
  2. Restructure: Move the charge into a downstream node that runs after the interrupt-bearing node returns approval into state. Now the interrupt-node only proposes; the next node executes.
  3. Idempotency belt-and-braces: Generate a charge_id in state before interrupt, pass it to the payment SDK as idempotency key — re-run becomes a no-op even if topology changes.
• 结果: One charge per approval; safe-by-construction.
• 可提取的操作: Two-line rule: (a) nothing irreversible before an interrupt() in the same node; (b) every external side-effect uses an idempotency key drawn from state. Sourced directly from the cheatsheet pitfall list [cheatsheet/gotchas].

反模式与边界 (Anti-patterns & Boundaries)

Concrete don'ts, each with the underlying reasoning.

• Don't graph a stateless pipeline. A 3-step prompt-tool-prompt chain with no cycles, no HITL, and no need to survive a crash does not need LangGraph. The abstraction overhead "could be a disadvantage in more straightforward scenarios" [duplocloud/compare]. Use a RunnableSequence.
• Don't use recursion_limit as a termination strategy. It "is not intended to be a primary control flow mechanism"; hitting it "indicates an underlying design flaw" [cheatsheet/gotchas]. Bake exit conditions into state.
• Don't put side effects before interrupt(). On resume, the node body re-runs from the top [cheatsheet/gotchas].
• Don't mutate state inputs. "Treat each node like a pure function: return a partial state update rather than mutating inputs" [swarnendu/best]. Mutation breaks checkpoint replayability.
• Don't share state between parallel branches without a reducer. Missing reducers on a key two nodes both write to triggers InvalidUpdateError [cheatsheet/gotchas].
• Don't run checkpointer.setup() at app boot. Treat it as a DB migration; run via CI/CD [bswen/hitl].
• Don't leave interrupted threads to rot. Implement a TTL sweep — without one, "state is held in the checkpointer indefinitely" [bswen/hitl].
• Don't use MessageGraph for new code. It's only "for basic chatbots"; every production case in this skill uses StateGraph [cheatsheet/gotchas].
• Don't parallelize blindly. "If one parallel node fails, the entire superstep fails atomically" — successful branches are discarded. Rate limits also hit faster [aipractitioner/scaling].
• Don't fan out with Send for fixed-cardinality work. Static parallel edges are simpler. Reserve Send for genuinely runtime-variable workloads [aipractitioner/scaling].
• Don't trust that an LLM-only loop will terminate. The text-to-SQL bug [gh/6731] is the canonical proof — always bound retries explicitly.

Hard boundaries (LangGraph is the wrong tool when):

• Latency budget < 200ms per call: checkpoint serialisation adds overhead.
• You need a visual drag-and-drop builder: that's Dify / n8n / LangFlow.
• You want zero-config role-based agents: that's CrewAI.
• You want a self-debugging code agent paradigm: AutoGen's strength [bswen/compare].

生态对照 (Ecosystem Context)

Source: LangChain's own production page [lc-built-with], the Bswen side-by-side comparison [bswen/compare], the OpenAgents comparison [openagents/2026], and the v1.0 vs functional-API blog [lc-blog/functional].

Decision heuristics:

• Reach for LangGraph when: durability matters, the workflow has cycles or long-running threads, multiple users share threads, you already use the LangChain ecosystem, or you need to insert humans into the loop without duct-tape [bswen/compare].
• Reach for CrewAI when: 24-hour proof-of-concept, role-based mental model fits the domain, no persistence needed. Many teams "use CrewAI for rapid prototyping to validate workflow logic, then port critical pipelines to LangGraph for production" [bswen/compare].
• Reach for AutoGen when: code-generation tasks where self-healing / iterative refinement is the point — but note Microsoft has shifted focus to the broader Agent Framework, so AutoGen is effectively in maintenance mode [bswen/compare].
• Reach for Swarm only when: studying multi-agent concepts as reference code. OpenAI itself labels it experimental [bswen/compare].
• Stay on plain LangChain when: no cycles, no state across turns, single LLM call or simple RAG. LangGraph is overkill [duplocloud/compare].

Internal LangGraph subdivision — also a choice point:

• create_react_agent (prebuilt): default for one tool-calling agent.
• Functional API (@entrypoint, @task): imperative Python style, shares the runtime, trades fine-grained time-travel for code brevity [lc-blog/functional].
• StateGraph: full control, required for multi-agent / parallel / custom routing.

Pick the smallest one that fits the requirements; promote upward as needed.

附录: 引用速查 (Citation Index)

Short tags used inline → full sources in references/:

• [lc-docs] = https://docs.langchain.com/oss/python/langgraph/*
• [lc-docs/why-langgraph] / [lc-docs/persistence] / [lc-docs/errors] / [lc-docs/streaming] / [lc-docs/subgraphs]
• [lc-blog/interrupt] = www.langchain.com/blog/making-it-easier-to-build-human-in-the-loop-agents-with-interrupt
• [lc-blog/benchmark] = www.langchain.com/blog/benchmarking-multi-agent-architectures
• [lc-blog/production] = www.langchain.com/blog/is-langgraph-used-in-production
• [lc-blog/functional] = www.langchain.com/blog/introducing-the-langgraph-functional-api
• [lc-built-with] = www.langchain.com/built-with-langgraph
• [gh/langgraph-readme] = github.com/langchain-ai/langgraph
• [gh/6731] = github.com/langchain-ai/langgraph/issues/6731
• [zenml/uber] = www.zenml.io/llmops-database/building-ai-developer-tools-using-langgraph-for-large-scale-software-development
• [alphabold/case] = www.alphabold.com/langgraph-agents-in-production/
• [bswen/hitl] = docs.bswen.com/blog/2026-04-16-langgraph-human-in-the-loop/
• [bswen/compare] = docs.bswen.com/blog/2026-04-29-agent-framework-production-comparison/
• [openagents/2026] = openagents.org/blog/posts/2026-02-23-open-source-ai-agent-frameworks-compared
• [eastondev/2026] = eastondev.com/blog/en/posts/ai/20260424-langgraph-agent-architecture
• [deepwiki/mapreduce] = deepwiki.com/langchain-ai/langchain-academy/7.1-map-reduce-pattern
• [deepwiki/subgraphs] = deepwiki.com/langchain-ai/langgraph/3.5-control-flow-primitives
• [swarnendu/best] = www.swarnendu.de/blog/langgraph-best-practices/
• [cheatsheet/gotchas] = sumanmichael.github.io/langgraph-cheatsheet/cheatsheet/faqs-gotchas/
• [bharatraj/state] = medium.com/@bharatraj1918/langgraph-state-management-part-1
• [duplocloud/compare] = duplocloud.com/blog/langchain-vs-langgraph/
• [dragonforest/timetravel] = dragonforest.in/time-travel-in-langgraph/
• [redis/checkpoint] = redis.io/blog/langgraph-redis-checkpoint-010/
• [aipractitioner/scaling] = aipractitioner.substack.com/p/scaling-langgraph-agents-parallelization
• [agentsindex/v1] = agentsindex.ai/blog/langgraph-tutorial