Mitosis vs LangGraph: choosing infrastructure for multi-agent systems#

If you are building with AI agents you eventually hit two separate problems: how the agent reasons, and what the agent knows. LangGraph is an answer to the first. Mitosis Cortex is an answer to the second. They are not alternatives, and choosing between them is usually the wrong question.

TL;DR#

• LangGraph is a Python and TypeScript library for expressing agent workflows as a stateful graph. You define the nodes and edges, and you run the graph yourself.
• Mitosis Cortex is a memory layer: it turns the email, chat, documents and records a team already has into one knowledge graph, and answers questions over it with citations.
• They solve different problems. A LangGraph workflow still has to get its facts from somewhere, and that somewhere is usually a pile of retrieved text. Cortex is the alternative to the pile.
• The common pattern is both: LangGraph decides what happens next, Cortex supplies the evidence each step reasons over.

What each one actually is#

LangGraph in one paragraph#

LangGraph is a graph-based agent runtime from the LangChain team. You define nodes (functions, LLM calls, tool calls), edges (transitions, often conditional), and shared state. The library executes the graph, persists state to a checkpointer, and exposes streaming, human-in-the-loop interrupts, and time-travel debugging. It is a library: you embed it in your app and run it on your own infrastructure.

Mitosis Cortex in one paragraph#

Cortex is a knowledge graph that belongs to one team. You connect sources, and it does the expensive work once while it ingests them: embedding each item, extracting typed entities and the relationships between them, and recording how each fact was derived. After that, any agent with access can query it. Retrieval fuses vector search, full-text search and a one-hop graph expansion, and returns scored, cited evidence. There is no model call at query time.

Side-by-side#

When LangGraph is the right call#

Pick LangGraph if:

• You are shaping the thinking of an agent and want explicit control over each step.
• Your workflow has branches, retries, or human approvals that benefit from a graph mental model.
• You want a library that drops into an existing app.
• You need deterministic replay of a run for debugging.

LangGraph is the right answer when the hard part is the shape of the reasoning.

When Cortex is the right call#

Pick Cortex if:

• Your agents keep being re-told the same things about your business, and you are paying for that re-explanation on every call.
• You need answers you can audit. Every Cortex result carries the source it came from, so a person can check it.
• Two spellings of the same customer, or the same project named three ways, are currently three different things to your retrieval layer.
• More than one agent needs the same picture of the business, and you do not want each one building its own.
• The cost of an answer matters. Cortex pays for the semantic work once at ingest rather than re-deriving meaning on every query, for every agent.

Cortex is the right answer when the hard part is what the agent knows.

Using them together#

This is the interesting case, and it is the common one:

1. Build the reasoning of your agent with LangGraph. Iterate locally.
2. Replace whatever retrieval step feeds your nodes with a Cortex query. It is one call, and it returns evidence rather than a pile of documents.
3. Cite what you used. Every result carries a universal id, so a node's output can point at what it was derived from.
4. Write conclusions back, so the next run starts from what the last one learned instead of rediscovering it.

You keep the parts of LangGraph you want (explicit reasoning graph, checkpointing, debugging) and stop hand-rolling the part Cortex already does.

Connect it through the MCP server if your agent speaks MCP, or through the TypeScript SDK if you would rather call it directly.

FAQ#

Is Cortex a LangGraph replacement?#

No, and it is not trying to be. LangGraph shapes how an agent reasons. Cortex is where the facts come from. Plenty of teams use both.

Can I use Cortex from a LangGraph node?#

Yes. A node can call the Mitosis SDK or the REST API directly, or reach the memory through the MCP server if your stack already speaks MCP.

What does "no LLM at query time" mean?#

Retrieval runs vector search, full-text search and a one-hop graph expansion, and fuses the results. No model is called to produce the answer. You get the evidence, and your model does the reasoning over it, which is why the retrieval step is measured in milliseconds and why a smaller model can often do the job.

How is this different from putting my documents in a vector store?#

A vector store finds text that looks similar to your question. It has no notion of identity, so the same person written two ways stays two things, and no notion of how a fact was established. Cortex normalizes entities so they collapse into one node, records typed relationships between them, and tags every fact with how it was derived, from a declared field at the trusted end down to a model's inference at the other. Retrieval can then rank by kind of evidence, not just by distance.

Does Cortex support MCP?#

Yes. Mitosis runs an MCP server, so a client that speaks the Model Context Protocol can read and write the memory without custom integration.

Where to go next#

• Read what Cortex is and why not files.
• Run the ingest and query loop end to end.
• See the full SDK surface if you want to know everything the client exposes.