jd five

Agent Frameworks

You just built refund workflows by hand:

  • one state object
  • nodes and edges
  • save / load for a human pause

That was on purpose. Now you can see what frameworks are for.

You do not have to keep writing the runner

A framework is a library that already implements common pieces:

Idea you learned What a framework often gives you
State object Typed state that merges updates
Nodes + edges Graph builder + runner
Branch on state Conditional edges
Save / resume Checkpoints (persistence)
Human pause Interrupts / approvals

You still design the story. You write less plumbing.

Same ideas. Less custom loop code.
  Your hand-rolled graph
    state + nodes + next() + JSON file

           │  same patterns
           ▼

  A framework (example: LangGraph)
    state + nodes + edges + checkpoint helpers

Same refund words → framework words

Our word LangGraph-style word
State State / channels
Node Node
Edge Edge
Branch Conditional edge
Save state Checkpoint / checkpointer
Human pause Interrupt / human-in-the-loop

When docs say “add a conditional edge,” you already know: look at state, pick next node.

What is common in 2026 (honest map)

Tool Strong for Languages (typical)
LangGraph State, graphs, checkpoints, HITL Python, TypeScript (official)
OpenAI Agents SDK Simple tool agents, handoffs Python, TypeScript
Claude Agent SDK Tools + MCP-heavy agents Mainly Python / TS ecosystem
Hand-rolled + provider SDK Any language, full control Python, Java, Go, TypeScript — what you used

There is no one framework that is first-class in Python, Java, Go, and TypeScript the same way. That is why this course taught patterns first.

If you use Python or TypeScript for production graphs, LangGraph is a common next step. If you use Java or Go, keep the hand-rolled pattern (or a community port) — you already understand it.

When to use a framework

Use a framework when… Stay hand-rolled when…
You need pause/resume often The job is a short tool loop
Many branches and long jobs You are learning the ideas (like now)
Your team already standardized on one No solid library for your language

Frameworks add versions, APIs, and opinions. They pay off when the plumbing would grow ugly.

What you should remember

  1. Frameworks wrap state, graphs, persistence, HITL — not magic new agent physics.
  2. Reading LangGraph (or similar) docs is easier after chapters 16–19.
  3. When a chapter needs a framework and we do not ship that SDK for a language, the Code panel stays empty — no complex stand-in.

Words to keep

Word Meaning
Framework Library that packages workflow plumbing (runner, checkpoints, …)
Checkpoint Framework word for a saved state snapshot
Conditional edge Framework word for a branch

See it in Code

The Code panel shows a small LangGraph refund graph in Python.
Other languages are empty here — the next chapter uses LangGraph.js / LangChain4j / Google ADK for a full agent.