Article Reader — Multi Agent

An n8n workflow that turns any article URL into a clean Arabic summary, using a planner → checker → writer → checker loop of four LLM agents that keep revising each other until the output passes review.

How it works

Chat trigger (URL)
  └─ HTTP Request          fetch the page
      └─ Extract Article Text   (CSS selector `body`, cleaned)
          └─ Planner        builds {title, main_idea, key_points[]}
              └─ Checker-P  reviews the plan  →  Plan OK?
                  ├─ no  → back to Planner with notes
                  └─ yes → Writer      writes the summary
                            └─ Checker-W reviews the draft → Draft OK?
                                ├─ no  → back to Writer with notes
                                └─ yes → Final Output

Every agent is paired with a Structured Output Parser, so each step emits strict JSON rather than free text. The checkers return {ok, notes}; when ok is false the notes are fed back into the previous agent and $runIndex is used to keep track of the attempt number.

Node	Role
Planner	extracts title, main idea, key points
Checker-P	reviews the plan, returns notes
Writer	writes the final Arabic summary
Checker-W	reviews the draft, returns notes
Final Output	shapes the response returned to the chat

All four agents run on Groq chat models (@n8n/n8n-nodes-langchain.lmChatGroq).

Requirements

• n8n (self-hosted or cloud) with the LangChain nodes installed
• A Groq API credential in n8n named Groq account

Install

1. In n8n: Workflows → Import from File
2. Pick Article_Reader_Multi_Agent.json
3. Open any of the four Groq Chat Model nodes and select your Groq credential
4. Activate, then open the chat and send an article URL

Notes

• Only the first 3000 characters of the article are sent to the Planner, to stay inside the context window. Raise that slice in the Planner prompt if your model allows it.
• The retry loops are bounded by the Plan OK? / Draft OK? IF nodes — if a checker never returns ok: true, the loop keeps running, so keep an eye on execution count while tuning prompts.

License

MIT