Forem: Merzouk Ayaden

I Ran 23 AI Agents 24/7 for 6 Months: Here's What Actually Broke (and How I Fixed It)

Merzouk Ayaden — Thu, 09 Apr 2026 23:53:04 +0000

I've been building and running a production multi-agent AI system for the past 6 months. 23 specialized agents, all running 24/7 on a self-hosted VPS (n8n + Claude + DeepSeek + Gemini).

This is not a tutorial. This is the honest breakdown of what actually broke in production, and how I fixed each problem.

The Setup

Platform: Self-hosted n8n (Docker, VPS)
Agents: 23 specialized agents (trading, content, monitoring, outreach, research, operations)
Models: Claude Sonnet/Haiku, GPT-4o/mini, DeepSeek Chat, Gemini Flash
Infrastructure: Traefik reverse proxy, Redis, PostgreSQL, Slack for alerts

Problem 1: API Costs Exploded

What happened: All agents defaulted to GPT-4o. Month 2 costs hit $180/month.

Root cause: No model routing. Every query — whether it was "format this JSON" or "analyze this 5000-word document" — went to the most expensive model.

The fix: A query classification layer before every LLM call.

// Complexity classifier (n8n Code node)
const query = $input.first().json.query;
const wordCount = query.split(' ').length;
const hasCode = /```
{% endraw %}
|function|class |import /.test(query);
const isComplex = hasCode || wordCount > 150;
const isMedium = wordCount > 50 && !isComplex;

if (isComplex) return [{ json: { model: 'claude-sonnet-4-5' } }];
if (isMedium) return [{ json: { model: 'claude-haiku-3-5' } }];
return [{ json: { model: 'deepseek-chat' } }];
{% raw %}

Result: 78% of queries classified as simple → DeepSeek at $0.001/query. New monthly cost: $22/month.

Problem 2: No Fallback = Total Outages

What happened: DeepSeek had a 2-hour outage. All agents using it stopped completely.

Root cause: Single provider, no fallback chain.

The fix: Primary → Secondary → Tertiary chain for every agent.


plaintext
DeepSeek (primary) → Gemini Flash (secondary) → Claude Haiku (tertiary)

One afternoon of work. 3 outages prevented since implementation.

Problem 3: Agents Got Stuck in Infinite Loops

What happened: An agent failed on step 3 of 7, retried indefinitely, consumed 40k tokens in 20 minutes.

Root cause: No max-attempt counter, no exit condition for failed states.

The fix:

Max attempt counter per task (configurable per agent)
Dead letter queue: failed tasks after max attempts go to a review channel
Alert on Slack for any task hitting the dead letter queue

Problem 4: Memory Was Fragile

What happened: After a VPS restart, agents lost all context. They had to "re-learn" their current state from scratch.

Root cause: Relying on in-context memory only. No persistent state.

The fix: PostgreSQL for all agent state. Every agent checks its state table on startup.


sql
CREATE TABLE agent_state (
  agent_id TEXT PRIMARY KEY,
  current_task JSONB,
  last_checkpoint TIMESTAMP,
  attempt_count INT DEFAULT 0
);

Agents now survive restarts cleanly.

Problem 5: Silent Failures Were Invisible

What happened: An agent stopped sending reports. Nobody noticed for 3 days.

Root cause: No monitoring. No alerts. No health checks.

The fix: A monitoring layer built into n8n:

Agent health check: cron every 5 minutes
Webhook availability check: every 10 minutes
Cost tracking: daily report with per-agent breakdown
Slack alert on any anomaly

First week after setup: Found 2 silent failures I had no idea existed.

What I'd Do Differently From Day 1

Design for failure first — assume every API, agent, and connection will fail
Add observability before adding agents — you can't fix what you can't see
Use a model router from the start — the cost savings pay for the implementation time in week 1
Define strict agent boundaries — scope creep in agent responsibilities creates unpredictable behavior
Build the dead letter queue early — failed tasks need a destination, not an infinite retry loop

Current State

After fixing all 5 problems:

Cost: $22/month (down from $180)
Uptime: 99.3% over last 30 days
Agents: 58 running (scaled up from 23 after costs were under control)
Monitoring: Full visibility, Slack alerts on any anomaly

Building multi-agent systems in production is nothing like the tutorials. The hard part isn't connecting the nodes — it's making the system resilient when things go wrong (and they will).

What's been the hardest part of your production agent setup? Happy to go deeper on any of the solutions above.

I Analyzed 3,000 AI Prompt Products on Gumroad — Here's What the Data Says

Merzouk Ayaden — Sun, 05 Apr 2026 22:33:48 +0000

I spent the last week analyzing the AI prompt market on Gumroad. Here's what I found — and why most builders are leaving money on the table.

The Market Reality

I scraped and analyzed 3,000+ AI prompt products across Gumroad. The data is brutally clear:

Price tiers and revenue distribution:

$10–19 range: 3,365 products = only 5.5% of total revenue (death zone)
$30–49 range: fewer products BUT 7.3% of revenue + higher avg sales (268 vs 241)
$50+ range: low volume, high margin, underserved

The biggest finding: ZERO direct competitors in the production multi-agent prompt niche.

Every other seller is offering:

Generic "ChatGPT prompts" (commoditized, $5–10)
Single-use marketing templates (high competition)
Midjourney image packs (unrelated vertical)

Nobody is selling a production-grade multi-agent operating layer with 49 battle-tested prompts from a live 35-agent system.

What the Top Sellers Do Differently

After studying the top 50 bestsellers on Gumroad:

Specificity wins: "57 prompts for your customer support agent" >> "100 ChatGPT prompts"
Social proof matters: "runs in production" beats "tested"
Price anchoring: $29 product with a $49 "pro version" converts better than standalone $19
Urgency through scarcity: limited time codes, countdown timers

The n8n + Multi-Agent Angle is Untapped

I searched Gumroad for "n8n", "multi-agent", "AI orchestration" — combined, fewer than 20 products in the entire marketplace.

Compare that to 15,000+ products for "ChatGPT prompts".

The demand side is different too: n8n has 400K+ GitHub stars, 50K+ community members. These are technical buyers who actually deploy AI agents. They pay more. They churn less.

What I Built Based on This Analysis

I run 35 autonomous AI agents in production across 65 n8n workflows. After this market research, I packaged our internal prompt library:

49 Production AI Agent Prompts — The Multi-Agent Operating Layer

What's inside:

49 prompts from a live production system (not generated, actually deployed)
10+ n8n workflow templates (lead gen, email automation, data sync)
5 Make.com blueprints ready to deploy
Guardian HUD setup guide (monitor all agents from one dashboard)
11,000+ automated tests passed

Priced at $29 (exits the "death zone" based on the market data above).

→ Get the kit here — use code LAUNCH49 for $10 off

Key Takeaways for Builders

If you're building AI automation products:

Get out of the $10–19 zone — 5.5% revenue share for 40% of products is not a business
Niche into specific agent types: support agent, sales agent, content agent
Lead with production proof: "10,000 tests" beats "battle-tested" beats "quality"
Build the price ladder: free sample → $29 core → $99 pro version
Underserved niches: n8n + AI agents + Make.com = almost no competition

The market data is clear: there's a window right now to own the technical AI agent niche on Gumroad before it gets crowded.

I post weekly about building AI automation systems in production. Follow for the next article on how we built a 57-agent orchestrator and what it actually took.

How I Automated 80% of My Freelance Workflow with n8n + AI Agents (4 Patterns That Actually Work in Production)

Merzouk Ayaden — Sun, 05 Apr 2026 19:43:48 +0000

I've been running n8n in production for 18+ months across dozens of client deployments. In this post, I'll share the 4 workflow patterns that appear in virtually every serious automation project — and how combining them with AI agents multiplies their impact.

Why Most n8n Tutorials Miss the Point

Most tutorials show you how to connect two apps. That's fine for demos — but real business automation requires something different: resilient, self-healing workflows that handle errors, adapt to data, and operate without constant supervision.

Here are the 4 patterns I use on every project.

Pattern 1: The Webhook → Validate → Branch → Act Loop

This is the backbone of any event-driven automation.

Webhook trigger
  → Data validation node (IF conditions)
  → Branch A: Happy path (process + notify)
  → Branch B: Error path (log + alert + retry)

Why it matters: Without validation, one malformed payload crashes your entire workflow. I always add a dedicated validation step that checks required fields, data types, and business rules before touching any downstream system.

Real example: Lead form submissions → validate email format + phone → if valid, push to CRM + send welcome sequence → if invalid, log to Airtable + alert Slack for manual review.

Pattern 2: AI Summarizer Injected in the Middle

This is where n8n becomes genuinely powerful in 2026.

Data input (webhook / sheet / email)
  → Normalize to clean text
  → OpenAI/Claude node (structured prompt)
  → Parse JSON response
  → Route based on AI decision

The key insight: Don't use AI as the endpoint — use it as a decision engine in the middle of your workflow. Ask it to classify, score, or extract structured data, then route accordingly.

Real example: Job application comes in → AI extracts skills, experience level, and culture fit score from CV text → if score > 80, auto-schedule interview → if 60-80, queue for human review → below 60, polite rejection email.

Pattern 3: The Async Queue with Slack Approval Gate

Some actions need a human in the loop — but that doesn't mean blocking your workflow.

Trigger event
  → Prepare action summary
  → Post to Slack with Approve/Reject buttons (Block Kit)
  → Wait for webhook callback
  → Execute or discard based on response

Why this rocks: You get automation speed with human oversight. The workflow pauses asynchronously — it doesn't block a thread or time out. The Slack message has all the context needed to make a decision in 10 seconds.

Real example: AI-generated content drafts → Slack message with preview + Approve/Edit/Reject buttons → on approval, auto-publish to website + social → on rejection, route back to editor queue.

Pattern 4: The Nightly Digest + Drive Sync

Every serious system needs a reporting layer.

Schedule trigger (nightly)
  → Pull data from multiple sources (CRM, DB, APIs)
  → Aggregate + format with AI
  → Save structured report to Google Drive
  → Post summary to Slack/Telegram
  → Optional: send email digest

Why it's essential: Without visibility, you're flying blind. This pattern gives you a daily snapshot of your system's health, revenue pipeline, and pending actions — without logging in to check five different dashboards.

Real example: Every night at midnight → pull Gumroad sales, Upwork proposals status, email open rates → AI formats into executive summary → saves to Drive → posts to team Slack channel.

Combining All 4: The Guardian Architecture

Here's how these 4 patterns combine in a real production system:

Incoming events hit Pattern 1 (validate + branch)
Complex decisions use Pattern 2 (AI in the middle)
High-stakes actions use Pattern 3 (Slack approval gate)
Everything is monitored via Pattern 4 (nightly digest)

The result: a system that processes hundreds of events per day, catches its own errors, escalates intelligently to humans, and gives you complete visibility — all running on a €20/month VPS.

Tools That Make This Work

n8n (self-hosted on VPS) — orchestration layer
OpenAI / Claude — AI decision nodes
Slack Block Kit — approval gates and alerts
Google Drive/Sheets — persistent storage and reporting
Airtable — structured data and queues
Webhooks — glue for everything

Get the Templates

If you want to skip the setup and start with production-ready versions of these 4 patterns, I packaged them (plus 96 more) in the Guardian AI Starter Kit:

👉 Guardian AI Starter Kit — N8N + Make.com + AI Agents

Includes:

100+ n8n workflow templates
All 4 patterns above (pre-built, documented)
AI agent configurations
Deployment guides for self-hosted n8n

Final Thought

The difference between a demo automation and a production automation is these 4 patterns. Demos show the happy path. Production handles everything else.

Build for the edge cases first. The happy path takes care of itself.

Building AI automation systems for clients? I'm available on Upwork — search for "n8n AI workflow automation" and look for Guardian AI systems.

I Built 65 n8n Workflows Running 24/7 With AI Agents — Here's What I Learned

Merzouk Ayaden — Sun, 05 Apr 2026 02:11:04 +0000

The Problem With Manual Workflows

After 3 months running a multi-agent AI system in production, I made a discovery: most automation guides show you how to build one workflow. Nobody talks about what happens when you have 65 of them running simultaneously.

Here's what I learned.

What's Running in My Stack

My Guardian AI system currently runs:

AI-powered lead generation pipelines — scraping, enriching, and routing leads automatically
Automated security scanning workflows — monitoring 10+ services, alerting on anomalies
Self-healing infrastructure monitors — workflows that detect failures and trigger recovery sequences
Sales tracking and CRM automation — pipeline stages updated without human touch
Multi-agent orchestration patterns — agents that spawn sub-agents and coordinate work

All tested with 11,000+ automated tests over 3 months of continuous operation.

3 Patterns That Actually Work at Scale

1. Checkpoint Everything

When you have 65 workflows, failures are inevitable. I use PostgreSQL checkpointing for every long-running workflow. If it crashes at step 47 of 50, it resumes from step 47.

2. Build Smoke Tests For Every Workflow

Every workflow has a corresponding smoke test. My suite checks all 10 public endpoints, all 7 Gumroad products, and all 8 Docker containers automatically.

Current score: 10/10 public endpoints OK.

3. Rate Limit Everything External

I built a rate-limiting layer into every workflow that touches an external API. This alone saved me from 3 separate API bans.

The Templates I'm Sharing

If you're building n8n automation:

n8n Workflow Templates Pack — 10 ready-to-deploy workflows (29 EUR)
Guardian AI Starter Kit — Full multi-agent system starter (29 EUR)
49 AI Agent System Prompts — Battle-tested prompts ($19)

What are you automating with n8n?