A save-game.
Every step of the agent is checkpointed per conversation thread. Crash, stop for a human, come back tomorrow: it picks up from the last save. Great for long, branching quests.
- theloop vs LangGraph
- 2-minute read
LangGraph remembers everything.
theloop remembers nothing.
Both run AI agents that call tools. LangGraph draws your agent as a flowchart and saves every step, so it can pause and pick up later. theloop takes one request, runs the loop and forgets. Your system keeps the memory.
In plain words.
An arcade machine.
One coin, one run, then a clean slate. Your system keeps the high score, which is the conversation history, and feeds it back in with the next coin.
Side by side.
| Question | LangGraph | theloop |
|---|---|---|
| What is it? | A framework for agents as graphs, in Python and JavaScript. Plus Agent Server, the server in LangSmith Deployment | A stateless HTTP service. Run it with Docker or Node |
| You describe the agent as… | Nodes and edges, in code | A JSON request: model, prompt, history, tools, limits |
| Where state lives | Checkpoints per thread. Agent Server keeps assistants, threads and runs in Postgres, plus Redis | In your database. The engine has no DB, no threads, no configs |
| Waiting for a human | interrupt() pauses the graph, and you resume it on the same thread |
A write tool answers deferred. You approve in your system, and the next run carries on |
| Server dies mid-run | Resumes from the last checkpoint LangGraph wins | The run is lost, and you're told so. Retry with the same history: no double writes |
| Many customers | Their threads and assistants live in its storage | Nothing to set up per customer: each request brings its own prompt, tools and key |
| Calling it | Python or JS SDK, or Agent Server's HTTP API | One POST /runs from anything that speaks HTTP |
| Best at | Long, branching, multi-agent workflows | One assistant loop inside your product, done carefully |
Which one is yours?
Pick LangGraph if…
- Your workflow has many steps, branches and loops, and you want to draw it
- A run must survive a crash and resume mid-step
- You orchestrate several agents that hand work to each other
- You're already on LangChain and LangSmith
Pick theloop if…
- Your product already has a database. One more place holding state is one too many
- You ship one assistant to many customers, each with its own prompt, tools and key
- Your backend is in any language, or it's a low-code platform
- You want the model to ask for actions, and your system to decide
The honest trade-off. theloop deliberately skips step-by-step durable execution. If a run dies, you start a new one with the same history. If that sounds scary for your workflow, LangGraph is the right tool.