advanced-ai-agent

Advanced AI Agent node for n8n with output validation (code checkers + AI validator), automatic retries with error feedback, required-tool-call enforcement, and structured JSON output via a formatter model

Package Information

Released: 3 mo ago
Downloads: 0 weekly / 0 monthly
Latest Version: 0.1.1

Available Nodes

Documentation generating…

Advanced AI Agent icon
Advanced AI Agent
AI Agent with output validation (code + AI validator), automatic retries with error feedback, and structured JSON output

Documentation

n8n-nodes-advanced-ai-agent

An Advanced AI Agent community node for n8n — a tools agent with a built-in quality-control loop:

  • ✅ Code Validation — run your own JavaScript checks on the agent's answer and its tool calls. On failure the agent is automatically retried with your error message as feedback.
  • 🤖 AI Validation — a second (or the same) chat model reviews the answer against a detailed, editable rubric. It sees the client message, the final answer, every tool call (name, arguments, result), and the agent's reasoning — so it can catch things like "you confirmed the order but never called the create_order tool". Rejections trigger a retry with the validator's feedback.
  • 📦 Structured Output — a formatter model converts the final validated answer into JSON matching an example you provide, with its own retry loop on parse/structure errors.
  • 🎛️ All three features are independent toggles. Extra model inputs (Validator / Formatter) only appear on the canvas when enabled.

How the retry loop works

input ──► Agent run ──► code check ──► AI validator ──► pass ──► formatter ──► output
              ▲              │               │
              └── feedback ◄─┴───── fail ────┘   (up to "Validation Retries" times)

On every failed check the agent is re-invoked with the conversation so far plus the failed answer and the feedback ("Your previous answer was rejected... you must call the create_order tool..."). Failed attempts are never saved to connected Memory — only the original input and the final answer are.

Installation

Self-hosted n8n → Settings → Community Nodes → Install → n8n-nodes-advanced-ai-agent.

Or manually: npm install n8n-nodes-advanced-ai-agent in your n8n installation's ~/.n8n/nodes directory.

Connections

Input Type When
Chat Model ai_languageModel always (required)
Memory ai_memory optional
Tool ai_tool optional, multiple
Validator Model ai_languageModel when AI Validation is on and Use Main Model as Validator is off
Formatter Model ai_languageModel when Structured Output is on and Use Main Model as Formatter is off

Code Validation contract

Your JavaScript gets these variables:

Variable Type Meaning
$input string the user message
$output string the agent's final answer
$toolCalls array [{ tool, args, result }, ...] — every tool the agent called
$reasoning string the agent's intermediate reasoning
$json object the incoming item's JSON

Return true (or nothing) to pass. Return an error-message string (or throw) to fail — that message is sent to the agent as retry feedback, so make it actionable:

if (/order|buy/i.test($input) && !$toolCalls.some((c) => c.tool === 'create_order')) {
  return 'The client asked to place an order but you never called the create_order tool. ' +
    'Call create_order with the requested items, then confirm the order to the client.';
}
return true;

Output

{
  "output": "final answer text",
  "validationPassed": true,
  "attempts": 2,
  "validationLog": [{ "attempt": 1, "source": "ai", "feedback": "..." }],
  "structuredOutput": { "reply": "...", "orderCreated": true },
  "toolCalls": [],
  "reasoning": ""
}

structuredOutput appears when the toggle is on; toolCalls/reasoning when Include Trace in Output is on. With On Validation Failure → Continue, exhausted retries emit the last answer with validationPassed: false instead of failing the node.

Notes

  • Requires self-hosted n8n (community nodes with external dependencies are not available on n8n Cloud).
  • The agent's "reasoning" is what LangChain captures between tool calls; providers' hidden native reasoning tokens are not always exposed by chat-model sub-nodes. Design validator rules to lean primarily on toolCalls + output.

License

MIT

Discussion