LangGraph for stateful agent workflows

How it works

LangGraph can checkpoint graph state and interrupt a run so an application can resume it or request a human decision. Edges make loops and branch conditions visible in the workflow.

What you need

Choose a state shape before adding nodes. If the application uses persistence, keep a stable execution identifier that can recover the intended run.

What to try first

Build a read-only lookup node and a human approval step, then stop the run and resume it using its saved state. Repeat the resume to check the effect on external calls.

Save the checkpoint immediately before a tool call. Resume from that state and verify the tool runs exactly as often as your application expects.

Where to look for failures

Durable workflow state does not make side effects idempotent. Design tool calls so a retried or resumed node cannot accidentally repeat a consequential action.

Related Franklin pages

Build the same two-step workflow with one tool and one failure case. Inspect how state, errors and tool permissions cross the boundary between the steps.

• Connect model-backed agents, tools and explicit workflows in Python or .NET. Microsoft Agent Framework for orchestrated agents
• Expose application functions to models through plugins and provider connectors. Semantic Kernel for integrating models into applications

Browse Franklin for concrete examples of tools and services exposed through MCP. Franklin’s MCP directory

Project documentation

• Current project README
• README at the reviewed revision