Review · Agent frameworks · August 2026
LangGraph
Best default runtime when the agent must pause, resume, and leave an artifact you can score.
8 / 10
Best for · Teams building stateful agents with human-in-the-loop writes, durable checkpoints, and explicit graph control.
Not for · People who want a named crew of roles in an afternoon and call the slide production.
TL;DR
LangGraph is the runtime I reach for when an agent must pause, resume, and survive a restart. Graphs, checkpointers, and interrupt() are the product. The OSS library is free. Hosted LangSmith Deployment and traces are where the bill shows up. Score the durable run, not the demo crew slide.
Built interrupt-and-resume loops with checkpointers; scored on whether the run survived a restart.
Product image
LangGraph product page preview — graphs and state, not freeform chat agents.
What it is
The graph.
Low-level agent orchestration library from LangChain. You define nodes, edges, and shared state. Checkpointers persist runs. interrupt() pauses before send, deploy, merge, or spend. Python and JS/TS. LangSmith covers traces, evals, and managed deployment.
Pricing
What it costs.
| Plan | Price | Note |
|---|---|---|
| LangGraph OSS | $0 | MIT-licensed library. Self-host. You pay LLM and infra bills. |
| LangSmith Developer | $0 / seat | About 5k base traces/mo included, then pay-as-you-go. One seat. |
| LangSmith Plus | $39 / seat / mo | About 10k base traces/mo, Deployment/Engine access, usage in LCU/LSU on top. Confirm on langchain.com/pricing. |
| LangSmith Enterprise | Custom | Hybrid or self-hosted options, SSO/RBAC, support SLA. |
Prices as of August 2026. Confirm on the vendor page — meters move.
Does well
What works.
Durable control flow
Cycles, retries, and typed state are first-class. A run can survive a restart if you wire a checkpointer. That is the difference between a chat toy and a production loop.
Human-in-the-loop that persists
interrupt() saves state and waits. Reviewers approve, edit, or reject later. Static interrupt_before flags are for debugging. Real gates use interrupt().
Traceable with LangSmith
When something fails at 2am, you need the graph path and tool args, not a vibes postmortem. Traces and evals sit next to the runtime.
Falls short
What fails.
You still design the agent
Low-level means you own the graph. Wrong edges, fat tool lists, and missing stop rules are your bugs. The framework will not invent a finish line.
Hosted bill is not the library price
OSS is free. LangSmith seats, traces, LCU/LSU, and deployments stack. Teams that treat Plus as a flat $39 get surprised.
LangChain lock-in
Deep LangChain coupling helps and locks. Escaping later costs more than starting thinner when the job is one tool loop.
Who should
Buy this if.
Builders who need durable agent runs with explicit grants and scored artifacts. Skip it if you only need a linear prompt chain or a weekend multi-agent costume party.
FAQ
Questions.
- Is LangGraph free?
- The library is MIT and free to self-host. Observability and managed deployment live in LangSmith and are metered.
- LangGraph or a single agent with tools?
- Start with one agent and a short tool list. Use LangGraph when you need persistence, interrupts, or multi-step control you can draw on one page.
- Does it replace LangChain?
- No. LangChain is integrations and components. LangGraph is the stateful runtime. They compose; they are not the same layer.
Also on the desk
Compare next.
Multi-agent frameworks · 6/10
CrewAI
Fastest way to stage a multi-agent demo. Still thin when grants, budgets, and a scored finish line matter.
Self-hosted automation / agents · 8/10
n8n
The automation runtime technical teams actually self-host: fair-code core, execution-based cloud pricing, and agents that share the same meter.
AI chat / Anthropic · 8/10
Claude
Best writing and careful reasoning in a chat box; Pro still runs out of air.