Gallery

Real agents doing real work, on the stage that suits each job. Every example is a small LangGraph agent you can read in one sitting, with the data it works on, so you can run it yourself and then swap in your own tools.

These aren't the keyless demos used elsewhere in the docs. Each one runs a real model, Claude Sonnet 5, unscripted, recorded in one take. The only editing is that long waits on the model are shortened; each page says by how much.

• 

  Support triage with approvals

  Web app · Board · human-in-the-loop

  A support agent works four tickets in parallel and waits for a manager to approve a $189 refund.

• 

  Sales analyst

  Web app · Python · inline charts · Canvas

  An analyst agent answers "what drove Q3?" from a year of sales data, with charts and a written report.

• 

  Incident investigator

  Terminal · log search · approval menu

  An on-call agent traces a spike of 503s to a bad deploy, fixes the config once you approve, and writes the incident note.

What they have in common

• Bring your own agent. Each one is a plain create_agent(...) graph with a handful of tools. LangStage doesn't ask you to build it a special way; you point a stage at it with --agent agent.py:graph.
• A human where it matters. Refunds over $100 and production config changes pause with LangGraph's interrupt(). LangStage shows the request and the decisions it allows in whichever stage you're using: the web app's dialog or Board, or the terminal's approval menu. See Human-in-the-loop.
• Real files. The agents read and write a workspace you can open: tickets and replies, a CSV and its charts, logs and an incident note.

The source for every example, and the scripts that recorded them, are in the docs repository under gallery/ and demos/gallery/.