First met in aicp-c3-langgraph-01 · Dev libraries & frameworks
Open-source framework for building an AI system as a graph instead of a line - so it can branch, run steps in parallel, go back and try again, and stop to ask a human.
The process has a decision, a retry that changes something, or a human pause. If it is a straight line, n8n is faster to build and easier to change.
Cycles - an edge that returns to an earlier node - are the reason it exists; a linear automation structurally cannot draw that arrow
Interrupts freeze a completed run for 5 seconds or 5 days at no cost, because state is checkpointed after every step
You draw the edges: the model reasons inside a node but never rewrites the map
The vocabulary is the skill - seven nouns are enough to specify a build to Claude Code or Codex without writing the Python
Holds one shared state record that every node reads and writes
Fans out independent nodes in parallel and merges their writes through a reducer
Routes on runtime conditions via conditional edges, including back into earlier nodes
Checkpoints state per step for cross-session memory, crash recovery and rewind
Pauses for human approval via interrupts before consequential actions
Doesn't suit linear pipelines - the trainer calls that over-engineering and points at n8n
Doesn't suit anything a non-technical ops team has to maintain
Doesn't trace itself - that is LangSmith's job
Doesn't cap its own cycles; an uncapped retry is a bill
The code-side answer when a workflow has to go backwards. Sits below an agent build and beside CrewAI, which solves the same problems differently.
Dev libraries & frameworks
No agent in the agency's stack currently needs a backwards arrow, but the retry-with-different-parameters shape is one Paul's audit work keeps approaching.
The state-as-one-record pattern is directly applicable to the KB extraction pipeline, where each session being extracted is effectively one job with many steps.
current
langchain: LangGraph is LangChain's graph framework; LangChain is the parent
n8n: explicitly the right choice for linear, trigger-driven work
LangSmith: LangGraph builds the agent, LangSmith records what it does
codex: the recommended way in: prompt in LangGraph vocabulary, let the coding agent write the Python
| Session | Coverage | Moment | What happened |
|---|---|---|---|
| Build Reliable AI Apps Using LangGraph: The Travel Agent That Changes Its Mind — State, Nodes, Cycles, Interrupts, and Why the Vocabulary Is the Skill | explained | The session's subject. Explained end to end - state, nodes, edges, parallel fan-out, reducers, conditional edges, cycles, checkpointers, interrupts - and prompted into Codex, but never actually running by the end of the session. Python chosen over LangGraph JS on air. | |
| Community Session 1: Causa Claims — a Non-Coder's Five-Agent Construction-Claims Pipeline, the Gate as the Product, and What RAG Really Costs | mentioned | Recommended next topic |