Mohamed Elghannam

Section 03 of 09

AI agents

Tools behind this sectionAI models and agents

Used to research, build and review, never to talk to customers. Each model has a job it is best at, and a person approves what leaves the system.

  • Claude and Claude Code

    What it does
    Language model and coding agent.
    How I use it
    Builds and tests the website, n8n workflows and this portfolio through MCP connections to n8n, GitHub, Cloudflare and the ad accounts; writes the Playwright checks.
    For a developer
    Engineering work that would need a developer is done and verified inside the marketing team.

  • Gemini

    What it does
    Multimodal model for text and images.
    How I use it
    Exterior renders at 2K, ad images with integrated Arabic and English text, and long-document reading through the API.
    For a developer
    Visuals for a new launch before the photography or the 3D studio is ready.

  • ChatGPT

    What it does
    General-purpose language model.
    How I use it
    Second opinion on copy and structure, and quick drafts that are then checked against the approved facts file.
    For a developer
    Faster drafting while the facts stay controlled in one place.

  • Perplexity

    What it does
    Answer engine with cited sources.
    How I use it
    One space per function (web and SEO, search ads, Meta analysis, video) for sourced research before a brief is written.
    For a developer
    Decisions backed by sources that can be checked, not by guesses.

How they work together

Research feeds the build; agents run inside n8n; a person approves anything customer-facing.ResearchBuildOperatePerplexitysourced researchChatGPTdrafts, second opinionClaude Codebuilds and testsGeminiimages, long documentsn8n agentscopilot, coach, intelManager approvesdrafts only, no auto-send

Research feeds the build; agents run inside n8n; a person approves anything customer-facing. Silent 10-second loop. Highlights follow the order the work actually happens.

03AI agents

Four n8n agents sit above the workflows for questions and analysis. Sensitive paths (intake, router, CAPI, funnel events) stay deterministic workflows. No agent is built or published without approval, and none talks to customers.

Copilot

Role
Answers the manager’s operational questions: yesterday’s CPL, which campaign is pacing, what the sales knowledge base says.
Tools
Read-only workflow over the sales knowledge base and verified market intel; data tables.
Limits
Single read tool, no write tools, observational memory off, verified rows only.

Coach

Role
Daily review of rep conversations: what was promised, what is missing, the next best message in a human tone.
Tools
SalesCoach review workflow, approved facts file, rounding rules for prices.
Limits
Banned claims stripped (escrow and bank-guarantee wording, “guaranteed return”). Output goes to the rep as a suggestion.

Intel

Role
Collects public market facts and competitor offers for Ajman.
Tools
Web search, save_market_intel_row.
Limits
Saves everything as pending_review; the manager verifies. No customer identifiers in searches.

Marketing

Role
Plans content jobs, shortlists and rotates creatives, audits coverage of the persona × angle matrix.
Tools
Content job table, creative-shortlist and matrix-coverage skills, sub-agent Coach.
Limits
Drafts only, publishing by the manager; approved facts only; rounded figures in ads, exact on price sheets.

Rule of thumb: workflows for anything that must be predictable; agents for questions with no fixed steps. Every agent version is recorded with a rollback ID.

n8n agents list with four published agents
Real workflow (sensitive values hidden) The four agents as published in n8n
n8n editor canvas of the website chat assistant workflow
Real workflow (sensitive values hidden) Site chat assistant workflow behind /api/chat

Want this for your projects?

The whole stack, or one piece of it. WhatsApp is fastest.

Message on WhatsAppBack to the summary