Week 3 assignment support: Build your automation or agent
This guide helps you ship the Week 3 assignment: a real automation or live agent for your product. Use the assignment brief for what counts as done; use this page to build it.
Time: 90 to 150 minutes for the general submission. Stretch goals are optional. Stuck more than 15 minutes is a WhatsApp question, not a solo struggle.
Tip
Session resources
Live patterns are in the Week 3 session deck. Hermes skills reference: official docs.
Step 0 — Pick the right shape
Before you open n8n or Hermes, decide who decides the steps. That is the whole Week 3 line.
Fixed workflow
SimplestYou know the path: when X happens, do A → B → C. Use n8n with normal nodes (no AI Agent). Best when the steps never change.
Single agent
n8n AI AgentYou know the goal and the tools, but not every step. One AI Agent node + tools (Sheets, HTTP, Slack). Best for "read this, decide, write that."
Multi-agent / orchestrator
StretchOne orchestrator delegates to specialists via AI Agent Tool. Use when one agent would be overloaded with conflicting jobs.
Always-on Hermes
ChannelYou want to message the agent on WhatsApp/Telegram and have it act with skills. Hermes is the front door; n8n can still be the hands via webhook.
Critical
Rule of thumb
Use the simplest thing that does the job. A plain workflow beats an agent when the path is known. An agent beats a brittle 20-branch IF tree when judgment is required.
Step 1 — Choose a product job (15 min)
Pick one job that is useful for *your* product or capstone — not a random demo. Examples:
- New signup / form submission → enrich or classify → write to Sheets or your DB → notify you
- Nightly schedule → summarise yesterday's activity → post a digest to Slack or WhatsApp
- Inbound support-ish message → draft a reply (human approves before send)
- Hermes on WhatsApp: "log this expense" / "what's my pipeline status" with 1–2 skills
Write one sentence: When [trigger], the system should [outcome] for [user]. If you cannot write that sentence, the job is still too vague.
Path A — Build in n8n (general submission)
- Create a new workflow. Start from a trigger: Webhook, Form, Chat, Schedule, or an app trigger you already use.
- If the path is fixed: add Do → Notify nodes only. Skip AI Agent.
- If you need judgment: add an AI Agent node. Attach a model + short system prompt (goal, constraints, what "done" looks like). Add one tool first (e.g. Google Sheets or HTTP Request).
- Run once with a known input. Open the execution log and find Think → Act → Observe. That is your proof it is an agent.
- Wire the output somewhere real: Sheets row, Slack message, email, or a webhook back to your Lovable/app.
- Activate the workflow so it runs without you clicking Execute.
Tip
n8n docs
First AI workflow: n8n intro tutorial. Templates: n8n.io/workflows — clone something close, then adapt.
Path B — Operate Hermes (general submission or stretch)
- Stand up Hermes and connect one channel (WhatsApp or Telegram).
- Write a short personality / instructions: who you are, what you will and will not do, how to ask for confirmation on irreversible actions.
- Add one skill with a clear "when to use" description. Test a message that should fire it and one that should not.
- Send a real task from your phone. Screenshot or record the reply and any tool/skill use.
- Optional: give Hermes a webhook skill that calls your n8n workflow — Hermes as front door, n8n as hands.
Step 2 — Prove it (required for general submission)
General submission evidence
- It runs on a trigger or schedule (not only manual Execute)
- It does something real for your product or a clearly named product job
- One line: workflow, single agent, multi-agent, or Hermes — and why
- Short clip / Loom / screenshots of it running end-to-end
- Export: n8n workflow JSON and/or Hermes config / skills screenshot
Stretch (optional)
- Qualification / scoring sub-agent or specialist
- Hot-path notification (Slack, email, or Hermes WhatsApp)
- Live Hermes with 2+ skills and one real task this week
- Graceful failure handling (empty results, bad tool calls, unknown senders)
- Connected to your Lovable / Pennywise / capstone product
- LinkedIn post (#BUILDINPUBLIC)
Step 3 — Stretch goals (how to level up)
- 1
Qualification / scoring agent
Add a specialist (or a second tool call) that scores Hot / Warm / Cold with a short reason, and write score + reason to Sheets or your DB.
- 2
Notifications
Only when score is Hot (or your equivalent), notify Slack/email — or have Hermes message you on WhatsApp.
- 3
Live Hermes
Channel connected + at least two skills + one real task completed this week. Optional: Hermes → n8n webhook.
- 4
Handle failures
Empty search, bad tool output, unauthorized WhatsApp sender — reply gracefully and log; do not crash or spam.
Common mistakes
Watch out
Avoid
- Building a clever agent when a three-node workflow would do
- Leaving the workflow inactive (only works when you click Execute)
- No product link — a generic calculator demo that is not yours
- Skipping the one-line classification (workflow vs agent)
- Stretch goals without a working general submission
Ready when
Ready to submit?
Post in Maven / WhatsApp with your clip, your one-liner (workflow vs agent), and which stretch goals you hit. Full brief: Add an automation or a live agent.