The AI Internship

Live session · 90 min

Agentic AI Builder's Bootcamp · Week 3

Agents, Multi-Agents
& Hermes.

From directing code to directing agents that run without you.

Act 1 · Foundations in n8n → Act 2 · Your live Hermes agent

Dr Aki Wijesundara

Instructor · Builder Bootcamp
Live build · agents → n8n → Hermes

Aki · The AI Internship01

The AI Internship

Today

Two acts. One idea: agents.

Act 1 · Foundations

Agents → multi-agents → n8n

What an agent actually is (goal, tools, loop)
Why one agent with ten tools fails
Orchestrator pattern in plain language
Build single-agent + orchestrator demos in n8n

Act 2 · Live agent

Hermes on your phone

Always-on assistant in WhatsApp / Telegram
Setup, personality, and skills
Cron / recurring jobs in natural language
Hand real work to it — then step back

Aki · The AI Internship02

The AI Internship

Bridge from Week 2

You already directed an agent.

Last week · Claude Code

Plan → act → you review

An agent in your files. You stayed in the loop at every approve step.

This week

Goal → tools → it decides

Agents that choose their own next step — and can run when you are not there.

Same posture: direct clearly, scope tightly. Different autonomy.

Aki · The AI Internship03

The AI Internship

Mental model

From directing code
to directing action.

Agents · Multi-agents · Always-on assistants

Aki · The AI Internship04

The AI Internship

Act 1 · Foundations

An agent is three things.

01

A goal — what success looks like, in plain language.

02

Tools — APIs, search, calculators, sub-agents it is allowed to call.

03

A loop — it thinks, acts, observes, and repeats until the goal is met or it stops.

Aki · The AI Internship05

The AI Internship

The agent loop

Not input → output. An autonomous loop.

Think

→

Act

→

Observe

→

Repeat

The model decides the control flow at runtime — which tool, in what order, when to stop.

Aki · The AI Internship06

The AI Internship

The line that matters

A workflow follows steps you defined.
An agent decides its own steps toward a goal.

Reach for the simplest thing that does the job

Aki · The AI Internship07

The AI Internship

Multi-agent theory

One agent + ten tools breaks.

Single agent overload

Picks wrong tools

Hallucinates parameters
Chains steps in the wrong order
Accuracy collapses as tools grow

Specialists

2–3 tools each

Clear responsibility boundaries
Failures stay isolated
Scale by adding agents, not tools

Aki · The AI Internship08

The AI Internship

Coordination patterns

Three shapes. We build #1 today.

Pattern 01

Orchestrator

One parent receives input and delegates to specialist agents as tools.

Best for chat + multi-step tasks

Pattern 02

Sequential

Agent A → B → C in a fixed order. Each passes work to the next.

Best for fixed pipelines

Pattern 03

Router

Classifier routes to the right agent. Deterministic, fast.

Best for high volume

Aki · The AI Internship09

The AI Internship

Orchestrator pattern

Parent decides. Specialists execute.

Orchestrator agentReceives the goal · picks who to call

Research

Web search · summarise

Structure

Parse · validate fields

Write

Save to Sheets / DB

Aki · The AI Internship10

The AI Internship

Act 1 · Live build

Now we build it in n8n.

The canvas where agent graphs become inspectable workflows.

n8n is the vehicle — agents are the idea

Aki · The AI Internship11

The AI Internship

Intro · n8n

Open-source automation with AI agents as first-class nodes.

01

Graph canvas

Triggers, actions, and AI nodes you can see and debug on one board.

Drag nodes, wire edges
Start from a template
Inspect every hop

02

Agent nodes

Chat model, memory, tools, and output parsers are built in.

AI Agent node
Window memory
Tool attachments

03

Agent Tool

One agent becomes a tool for another — the multi-agent primitive.

Orchestrator parent
Specialist children
Runtime delegation

04

Inspectable

Every run is logged — see what the agent decided and why.

Input / output per node
Tool call traces
Debug without guessing

Aki · The AI Internship12

The AI Internship

Building blocks

Four node families. Focus on AI.

Triggers

Entry points — nothing runs until one fires.

WebhookChatScheduleForm

Actions

Talk to the outside world.

HTTPSheetsGmailSlack

Logic

Control flow and reshape data.

IFSwitchSetCode

AI / LangChain

Where agents live on the canvas.

AI AgentMemoryToolsAI Agent Tool

Aki · The AI Internship13

The AI Internship

Live demo 1

Single agent + tools.

Chat trigger → AI Agent → calculator (or a product tool). Watch the loop in the execution log.

Feel the difference: guess vs tool call

Aki · The AI Internship14

The AI Internship

Demo 1 · Architecture

One agent. One tool. Full loop.

Chat Trigger

↓

AI Agent

↓ tools

Calculator · Chat Model · Memory

Why it matters

Without the tool, the model guesses. With it, the agent delegates to a deterministic system — then continues the loop.

Build checklist

Chat trigger → Agent → Model → attach tool → test → open execution log.

Aki · The AI Internship15

The AI Internship

Live demo 2

Orchestrator + specialists.

Type a name → research agent → structure agent → write to Sheets. The parent decides the order at runtime.

Multi-agent · AI Agent Tool nodes

Aki · The AI Internship16

The AI Internship

Demo 2 · Architecture

Lead enrich & route — delegated.

Orchestrator“Infer the lead, research, structure, save — don’t ask me”

Deep Research

Web search + structured output

Sheets Writer

Append validated row

Several agent iterations. Order decided at runtime — we did not hardcode the path.

Aki · The AI Internship17

The AI Internship

Act 2

n8n built the graph.
Hermes makes it yours.

Always-on · Messages you first · Remembers

Aki · The AI Internship18

The AI Internship

What Hermes actually is

An always-on agent in your channels.

01

Persistent — runs on your server, keeps memory across conversations.

02

Channel-native — WhatsApp, Telegram, Slack, Discord, CLI from one gateway.

03

Skills + cron — packaged abilities and scheduled jobs, not just chat replies.

Not a model provider. An open-source assistant you operate (Nous Research Hermes Agent).

Aki · The AI Internship19

The AI Internship

Mental model

Front door · brain · hands.

Front door

WhatsApp / TelegramWhere you already live. The channel is not the intelligence.

Brain

HermesPersonality, memory, skills, cron, and the decision to act.

Hands

n8n (stretch)Known steps and multi-agent graphs — wire via webhook/skill when you need heavy workflows.

Aki · The AI Internship20

The AI Internship

Live build · Hermes

Stand up your agent in four moves.

01

Install Hermes

VPS or local — follow the setup sheet, or hermes setup from the docs.

02

Choose your model

hermes model — Claude, OpenAI, OpenRouter, or Nous Portal. Add your API key.

03

Connect a channel

hermes gateway setup — Telegram is fastest; WhatsApp via the built-in bridge.

04

Message it for real

Start the gateway, send a test, confirm it replies with memory and tools live.

Aki · The AI Internship21

The AI Internship

Make it yours

Give it a personality and a job.

Starter instructions · copy & adapt

You are my product co-pilot for Pennywise (personal expense tracker).
Be concise on WhatsApp. Prefer action over essays.
When I send an expense or a budget question: clarify only if critical, then act.
Never spend money or email customers without asking me first.
Use skills I install. Prefer the simplest path that works.

Aki · The AI Internship22

The AI Internship

Skills

Skills are packaged abilities.

Author

Name · when · what

SKILL.md-style descriptions (agentskills.io). Clear enough that the agent knows when to reach for it — same craft as Week 2 skills.

When to use this skill
What it does step by step
Test: fires when it should, skips when it shouldn’t

Install carefully

Skills Hub ≠ trusted

Prefer skills you wrote or reviewed. Treat third-party skills like installing plugins — audit before you attach them to a live channel.

Start with 1–2 skills for real work
Hermes can also create skills from experience
Scope access on WhatsApp / Telegram

Aki · The AI Internship23

The AI Internship

Always-on work

Then give it a real recurring job.

Example · daily brief · adapt to your product

Every weekday at 9:00, research 5 high-fit leads for my AI product.
For each: company, decision-maker, role, why they fit, site, LinkedIn if found,
and one personalised outreach angle. Send the results to this WhatsApp chat.
Split into multiple messages if needed.

Aki · The AI Internship24

The AI Internship

Consolidate

n8n, or Hermes?

	n8n	Hermes
When	Steps known, or multi-agent graphs	Ongoing assistant in your channel
You give	Nodes, tools, orchestrator prompts	Personality, skills, channel
You get	Inspectable runs & integrations	Always-on presence + memory
Together	Heavy graphs, inspectable runs	Stretch: call n8n via webhook/skill

Aki · The AI Internship25

The AI Internship

Where builders trip

Four failure modes — and the fix.

Reached for an agent when a workflow would do → simplest thing that works.

One mega-agent with ten tools → specialists + orchestrator.

Unreviewed community skills → author or audit before install.

Vague goal, no confirmation rules → scope like a Week 1 brief.

Aki · The AI Internship26

The AI Internship

Assignment

Ship a live agent — or a multi-agent flow.

Pick one

Hermes instance on your channel + 2 skills + one real task this week
n8n orchestrator (or single agent + tools) wired to your product
Stretch: Hermes calling an n8n webhook (front door → hands)

Done when

Something useful runs without you clicking each time
You can say: workflow, single agent, or multi-agent — and why
Short clip or Loom + config screenshot in WhatsApp

Full brief on the syllabus →

Aki · The AI Internship27

The AI Internship

Wrap

#BUILDINPUBLIC

Post a short clip of your agent acting — and one line: workflow, single agent, or multi-agent.

Aki · The AI Internship28

The AI Internship

Week 4 preview

Evals: the TRACE loop.

“It looked good when I tried it” is not evaluation.

Shared session with the Engineering cohort

Aki · The AI Internship29

The AI Internship

Questions in the WhatsApp group.

Recording and slides land after the session.

See you in Week 4

Aki · The AI Internship30

Agents, Multi-Agents& Hermes.

Two acts. One idea: agents.

Agents → multi-agents → n8n

Hermes on your phone

You already directed an agent.

An agent is three things.

Not input → output. An autonomous loop.

One agent + ten tools breaks.

Three shapes. We build #1 today.

Orchestrator

Sequential

Router

Parent decides. Specialists execute.

Open-source automation with AI agents as first-class nodes.

Four node families. Focus on AI.

Triggers

Actions

Logic

AI / LangChain

One agent. One tool. Full loop.

Lead enrich & route — delegated.

An always-on agent in your channels.

Front door · brain · hands.

Stand up your agent in four moves.

Give it a personality and a job.

Skills are packaged abilities.

Then give it a real recurring job.

n8n, or Hermes?

Four failure modes — and the fix.

Ship a live agent — or a multi-agent flow.

Agents, Multi-Agents
& Hermes.