How can I chain local LLaMA-based AI agents in n8n to sequentially build a complete article?

Below is some guidance on making your agents build on previous outputs in your n8n workflow when using local LLaMA models:

---

Aggregating Outputs Between Agents

By default, if you simply pass the output of one agent directly into the next, the new output may replace the previous output. To “accumulate” the article parts, you’ll need to explicitly merge or append the generated texts. Consider these options:

• Use Merge or Set Nodes:
  Use a Merge node (or a combination of Set nodes) to aggregate all the parts from the different agents. For example, after each agent generates its section (headline, summary, paragraphs, conclusion), use a Merge node with the “Combine” or “Append” option to concatenate the texts.

• Pass Prior Content as Part of the Prompt:
  When calling the next agent, include the text generated so far in its prompt. You can do this by configuring the input prompt with an expression referencing the output of previous nodes. For instance:

  JAVASCRIPTCopied
  "Generate a summary for the following article so far:\n\n{{ $json[\"accumulatedArticle\"] }}\n\nInstructions: ..."

  This way, the agent sees the full context and can add its section without overwriting earlier content.

---

Techniques Specific to Local LLaMA Implementations

While cloud-based services (like OpenAI) may have built-in conversation memory, for local LLaMA models you might have to implement context management yourself. In n8n this can be achieved by:

• Customizing the Workflow Prompts:
  Ensure that every time you call a local LLaMA agent, the input prompt includes all previous sections of the article. This requires storing the output in a variable (using a Set node or similar) and then referencing that variable when preparing the prompt for the next agent.

• Using Memory Sub-nodes (If Available):
  If your implementation of the AI Agent node supports memory (for example, a “Window Buffer Memory”), connect and configure it so that it retains conversation history. Verify that the memory configuration passes along previous text to each subsequent agent call. Check the local model’s documentation or your node’s configuration options to ensure that memory isn’t reset between runs.

• Ensuring Correct Parameter Passing:
  In your environment (such as using a Qdrant Vector Store with SQLite), make sure that any settings related to statefulness (or memory retention) are appropriately configured. Although this is more relevant for cloud services, your local implementation may also require explicit parameters to “append” text rather than overwrite.

---

Example Workflow Steps

1. Agent 1 (Topic & Instructions):

   • Receives a topic suggestion and creates an initial piece of text with instructions.
   • Store this output in a variable (e.g., “accumulatedArticle”).

2. Agent 2 (Headline):

   • Uses an expression in its prompt like:

     "Based on the following instructions, generate a headline:\n\n{{ $json.accumulatedArticle }}"
     

   • Append the headline to your stored variable using a Merge or Set node.

3. Subsequent Agents (Summary, Paragraphs, Conclusion):

   • Each agent’s prompt includes the previously merged article parts (e.g., “accumulatedArticle”), ensuring that they build on the complete context rather than replacing it.
   • After each agent finishes, use a Merge node to append the output to the overall article.

---

Final Recommendations

• Design Your Prompts Carefully: Always include previous outputs in the prompt for the next agent.
• Test with Simple Examples First: Begin with just two agents (e.g., one generating a headline, the next a summary) to verify piping and merging of text.
• Review Your Node’s Configuration: Confirm that no node is inadvertently set to clear or overwrite stored data.

By explicitly managing how data is passed and merged between agents, you can create a collaborative workflow where each agent contributes a part of the article rather than erasing what came before.

Feel free to ask if you need further details on setting up specific nodes or expressions!