Harrison Chase on Production Agents #2: LangGraph as the runtime

Part 2 of 5 — tracing Chase's production agent thinking with FRE|Nxt Labs commentary.

---

The announcement

"🚨 New DeepLearning course: AI agents in LangGraph. LangGraph is an extension of LangChain for building agent & multi-agent systems. I believe it's the best way to build agents for a few reasons: 🔀 Controllability 🧠 Memory 🫂 Novel UX paradigms"

— Harrison Chase (@hwchase17), June 2024

Later framing from a 2025 interview: "LangGraph is the runtime. LangChain is the abstraction. Deep Agents are the harness."

---

What we heard

LangChain (entry #1) solved the composition problem: how do you wire an LLM to tools and prompts? LangGraph solves a different problem: how do you run the wire?

The shift is from library to runtime. A library gives you functions to call. A runtime manages state between calls, persists it, resumes it, lets you time-travel through it. That difference is the entire gap between "my agent works in a notebook" and "my agent works at 2 AM on Tuesday for a customer I'll never see."

Three specific things LangGraph provides that a plain LangChain composition doesn't:

1. State as a first-class object: the graph has state; every node reads from it and writes back to it. You don't pass context through nested function arguments.
2. Checkpointing: state is persisted at each node automatically. Failed sessions resume from the last successful node.
3. Human-in-the-loop as a node type: pausing for approval is a structural primitive, not a bolted-on callback.

If your agent doesn't need those three things, you don't need LangGraph. If it does, you do.

---

What we actually do with this

We treat LangGraph as the default runtime for any PRL-3 or higher agent system (see Amodei #1 for the PRL framework). Our standard LangGraph architecture per engagement:

| Graph component | Used for | |---|---| | State schema | The single source of truth for session state. All node inputs and outputs flow through it. | | Node per agent role | One LangGraph node per persona/capability. Each node is small, testable, and independently versioned. | | Edges as routing | Conditional edges encode the state machine. No if/else in agent code — all branching lives in edges. | | Interrupts for approval gates | Every human-in-the-loop checkpoint is a LangGraph interrupt. Resumable from exactly that point. | | Checkpointer backed by Postgres | Session state survives restarts. Replays are free. |

Rule we enforce: no stateful agent without a graph. The moment a system needs memory across turns, it runs on LangGraph. Chat wrappers, retrieval pipelines, and tool-calling loops can live elsewhere. Multi-turn agents with state cannot.

---

Applied: InterviewLM's graph

InterviewLM is one LangGraph per candidate session. The graph has 12 nodes:

• 1 entry node (session setup)
• 1 coordinator node (routes between interviewer personas)
• 5 interviewer persona nodes (technical, behavioural, case study, closing, follow-up)
• 3 evaluation nodes (score aggregation, rubric matching, recommendation)
• 1 human-in-the-loop node (optional live interviewer intervention)
• 1 exit node (session summary + transcript storage)

Every node reads and writes to a shared SessionState schema. The checkpointer writes to Postgres after each node. A failed session — for any reason, including LLM provider outage — resumes from the last successful checkpoint. We've had zero data loss on 100+ concurrent sessions over the engagement lifetime.

The design cost: roughly one week of state-schema design upfront. The design payoff: every failure scenario has a deterministic recovery path. We never have to reconstruct "what was the agent doing when it died?" — the state tells us.

---

The one thing to steal from this

On your next agent system, model the state before writing a single node. Draw the state schema on a whiteboard. Every field must be justified: why does this field exist, who writes it, who reads it? If you can't defend a field, it shouldn't be in the schema. The state design is the architecture. Everything else is implementation detail.

---

Next in this series

#3 — Better models alone won't ship your agent (2025). Chase's central argument in every interview and keynote since 2024: the gap between prototype and production is orchestration, not model capability. What we run before calling an agent production-ready.