n8n-nodes-aot-harness

CHIP + Atom of Thoughts (AoT) for n8n — multi-provider agent harness with built-in cost tracking, plus a Process Analyst Worker for operational case triage.

A community node that brings the AoT-Harness (atomic decomposition + QA loop) into n8n.
One node, one goal — the harness decomposes the task into atoms, runs them in parallel, QA-scores the result, and returns the polished output plus a per-call cost breakdown.

What's new in v0.5.0 — Process Analyst Worker

The Process Analyst Worker turns unstructured operational input into structured next-step decisions for n8n workflows: classification, missing information, urgency, automation potential, risk flags, human review and draft response.

Der Process Analyst Worker übersetzt unstrukturierte operative Eingaben in strukturierte nächste Schritte für n8n-Workflows: Klassifikation, fehlende Informationen, Dringlichkeit, Automatisierungspotenzial, Risiken, Human Review und Antwortentwurf.

• 🆕 New mode Process Analyst — single-call structured triage of one operational case (email, document, support ticket, workshop request, internal note)
• 🧠 4 process profiles — generic, email_support, document_intake, automotive_service
• 🔒 Read-only by design — produces analysis + draft response, never sends. Human-in-the-Loop is part of the contract.
• ⚙️ Deterministic HITL gate — human_review_required is forced true whenever confidence falls below the threshold, missing facts exist, risk flags appear, or input is too short to be reliable
• 📦 Structured JSON output — drop directly into Switch / Set / Gmail / Supabase nodes downstream

See the Process Analyst Worker section below.

What's new in v0.3.0 — Multi-Provider

• 🌐 5 providers: Anthropic, OpenAI, Google Gemini, Mistral (EU/GDPR), OpenRouter (100+ models, 1 key)
• 🪙 Mixed-Provider Mode — let a smart model decompose, a cheap one execute. Typical saving: 60–80% at comparable quality.
• 💰 Cost tracking in node output — every run reports cost.total_usd, tokens, breakdown by provider and by model
• 🇪🇺 Mistral (la Plateforme) — fully EU-hosted, GDPR-friendly, ideal for German/EU SMB workflows

⚠️ Breaking change for existing v0.2.x users — see Migration below.

Install

In n8n → Settings → Community Nodes → Install:

n8n-nodes-aot-harness

Then add the credential(s) for the provider(s) you want to use:
AoT Harness — Anthropic / OpenAI / Google Gemini / Mistral / OpenRouter.

Quick Start

1. Drop AoT Harness into a workflow.
2. Pick a Provider (e.g. Anthropic) and a Model (e.g. claude-sonnet-4-6).
3. Attach the matching credential.
4. Set the Goal field to the task in plain language:

   Erstelle eine kurze IDD-Dokumentation für einen Kunden, 42 J., sucht private Haftpflicht…
   

5. Run.

The node returns:

{
  "goal":      "...",
  "result":    "polished final output",
  "qa_score":  0.92,
  "success":   true,
  "atoms_done": 4,
  "atoms_total": 4,
  "provider_used": "anthropic",
  "model_used":    "claude-sonnet-4-6",
  "cost": {
    "total_usd":         0.0143,
    "total_calls":       6,
    "prompt_tokens":     2480,
    "completion_tokens": 1120,
    "by_provider":       { "anthropic": { "cost_usd": 0.0143, "calls": 6, ... } },
    "by_model":          { "claude-sonnet-4-6": { "cost_usd": 0.0143, ... } }
  }
}

🪙 Mixed-Provider Mode (cost-saver)

Toggle "Enable Mixed-Provider Mode" in the node:

• Decomposer = the smart model that splits the goal into atoms (e.g. Claude Opus / Sonnet)
• Executor = the cheap model that solves each atom and runs QA (e.g. Gemini Flash)

Typical results on a German IDD-documentation task:

Annualized at 1k runs/month: ~$130/year saved per workflow.

A ready-to-import demo workflow ships in examples/mixed-provider-cost-saver.json — runs both setups against the same goal and reports the delta in a Code node.

🛡️ Production demo: Insurance-broker claim triage

A full Versicherungsmakler-use-case lives in examples/schadenmeldung-triage/:

• Goal: incoming claim email → 2 ready-to-send Gmail drafts (to customer + to insurer) in ~60s
• Features shown: AoT decomposition (4 atoms), Mixed-Provider cost saving, QA gate with HITL fallback (Makler gets review email if qa_score < 0.75)
• Economics: ~20.000€/year savings for a broker handling 50 claims/month — full math in the folder's README.md
• Includes: importable workflow, test mail, Supabase seed SQL for automatic policy lookup

This is the workflow to show an interested broker.

Providers & default models

Per-provider model dropdowns are pre-curated in the node UI.

How it works

Goal
  │
  ▼
[Decomposer LLM]   AoT decomposition → AtomGraph (1–6 atoms, dependency-aware)
  │
  ▼
[Executor LLM]     Atoms run in parallel via Promise.all (where dependencies allow)
  │
  ▼
[Executor LLM]     QA-Agent scores 0–1 (retry on failure)
  │
  ▼
{ result, qa_score, success, atoms_used, cost }

In single-provider mode Decomposer = Executor.
In mixed mode they're independent — different provider, different model, different credential.

Modes

🛠️ Process Analyst Worker

A focused mode for operational case triage. Pick Mode → Process Analyst, paste one inbound case (email, support ticket, document excerpt, workshop request, internal note), and the worker returns a single structured JSON payload that the rest of the n8n workflow can branch on.

When to use it

• Inbound email needs to be classified, scored, and possibly answered (or escalated)
• A scanned document arrives and you need to know: which type, what's missing, what to do next
• A workshop / car-dealership receives a service request and you want a structured intake before a human picks it up
• Any process where an unstructured input must become a deterministic next step

Parameters

The active Provider + Model at the top of the node is reused as the analyst LLM (no separate decomposer needed — this is a single call).

Example: automotive service request

Operational Input:

Sehr geehrte Damen und Herren,

mein Golf macht beim Bremsen seit ein paar Tagen ein schleifendes Geräusch.
Können Sie mir nächste Woche einen Termin anbieten?

Viele Grüße
Max Mustermann

Profile: automotive_service · Language: de · Threshold: 0.7

Output:

{
  "mode":    "process_analyst",
  "profile": "automotive_service",
  "analysis": {
    "process_type":  "terminanfrage",
    "business_area": "service",
    "summary":       "Kunde fragt nach einem Werkstatttermin für nächste Woche wegen Bremsgeräuschen am VW Golf.",
    "detected_entities": {
      "customer_name":   "Max Mustermann",
      "company":         null,
      "email":           null,
      "phone":           null,
      "vehicle":         "VW Golf",
      "license_plate":   null,
      "requested_date":  "nächste Woche",
      "topic":           "Bremsgeräusch / Werkstatttermin"
    },
    "urgency":     "medium",
    "confidence":  0.82,
    "missing_information": [
      "Telefonnummer",
      "Kennzeichen",
      "konkreter Wunschtermin"
    ],
    "next_best_action":      "Rückfrage an Kunden mit Bitte um Kennzeichen, Telefonnummer und zwei mögliche Zeitfenster für nächste Woche vorbereiten.",
    "automation_potential":  "high",
    "human_review_required": true,
    "risk_flags": [
      "Terminwunsch unklar",
      "Fahrzeugdaten unvollständig"
    ],
    "recommended_workflow_tags": [
      "service_intake",
      "appointment_request",
      "human_review"
    ],
    "draft_response": "Hallo Herr Mustermann, vielen Dank für Ihre Anfrage. Damit wir Ihnen einen passenden Werkstatttermin anbieten können, senden Sie uns bitte noch Ihr Kennzeichen, eine Telefonnummer und zwei mögliche Zeitfenster für nächste Woche.",
    "_profile":         "automotive_service",
    "_trigger_reasons": ["3 missing fact(s)", "2 risk flag(s)"]
  },
  "provider_used": "anthropic",
  "model_used":    "claude-sonnet-4-6",
  "cost": { "total_usd": 0.0021, "total_calls": 1, "prompt_tokens": 480, "completion_tokens": 220 }
}

The downstream Switch node branches on analysis.human_review_required → either send the draft via Gmail or route to a Makler/Service-Annahme inbox.

Logic guarantees

• Read-only. No mailing, no DB writes, no decisions on legal / monetary topics.
• human_review_required is deterministic — set in code (not just by the model) whenever any of these is true:
  • confidence < humanReviewThreshold
  • missing_information.length > 0
  • risk_flags.length > 0
  • input is shorter than 20 characters
• _trigger_reasons lists exactly which of the above fired — useful for explainability + debugging.
• recommended_workflow_tags always contains human_review when the gate fires, so a Switch node can match on a single tag.

Profile cheatsheet

Migration from v0.2.x

v0.3.0 is a breaking change because the node now requires a provider-specific credential.

After updating:

1. Open every workflow that uses AoT Harness.
2. Set Provider → Anthropic (matches the v0.2.x default behavior).
3. Attach the new AoT Harness — Anthropic credential. Re-enter your ANTHROPIC_API_KEY.
4. Pick a Model from the dropdown (default: claude-sonnet-4-6).
5. Save & test.

The legacy aotHarnessApi credential type is still registered so existing credential entries don't disappear from your n8n credentials list — but the new node doesn't read it. Delete it after migration if you like.

Credentials

You only need credentials for providers you actually use.

Roadmap

• v0.5.1 — Process Analyst: per-profile entity-schema overrides, optional batch-mode for multi-case analysis
• v0.6 — additional profiles (legal_intake, finance_intake, hr_intake), webhook-mode parity for the analyst worker
• v0.3.1 — per-atom provider override (route specific atoms to specific models)
• v0.4 — Mistral self-hosted via Ollama, ReAct-style tool loops, prompt-cache visibility for non-Anthropic providers

Based on

• AoT Paper: arXiv:2502.12018 (NeurIPS 2025) — MIT License
• CHIP Architecture: Ronny Schumann
• Harness Concept: Anthropic Engineering, Martin Fowler (2026)

License

MIT