Agentic AI for Applied Learning

A one-day course ↯

Building AI Agents
without losing your mind.

Learn agent principles, scale them, and design real-world patterns

principles

How We'll Work Together

  • Default to curiosity
    There's usually something useful in every idea.
  • Bias toward action
    Don't overthink. Try it, test it, and learn by doing.
  • Make thinking tangible
    Externalise your ideas. Sketch, type or say it out, or build on others so we can think together.
today's journey

From n00b to AI Agent Strategist.

n00b AI Agent Strategist
Foundations
What agents actually are
Anatomy
Inside an agent
Design
Agentic design patterns
Evaluations
Evaluate your agent
First Build
Your first agent
Multi-Agent
Level up your agents
Workshop Session
Solve for a real client
↯ click the cards ↯ why you're here today

The Agentic Opportunity

"

For every SaaS company, there will be an AI agent company.

- Harj Taggar, Partner Y Combinator

flip to reveal ↗

Agents democratize economic opportunities.

"

Fast experimentation drives invention.

- Andrew Ng, Computer Scientist, former head of Google Brain & Chief Scientist at Baidu

flip to reveal ↗

The key enabler of AI is fast development and iteration, and that allows for more experimentation.

proof point · real-world deployment

SingTel Case Study

0 wks
Went live in under 10 weeks with Sierra's forward-deployed engineer (FDE)
0+
Customer cases handled in the first six weeks
0%
Mobile and home troubleshooting resolved without a human agent
0%
Roaming sign-ups completed without human intervention
freed
up
Customer care officers freed up to focus on complex, higher-value work
↯ click to explore ↯ the definition

What is an agent?

A model that perceives inputs, reasons about what to do, takes actions - and keeps going, step by step, until the goal is met.

GOAL PERCEIVE text · files · data · web REASON plan what to do next ACT use tools · take steps ITERATE check · retry · adapt
Core elements
Brain The LLM core

The LLM enables the agent's reasoning. You swap models to change its reasoning engine. The agent is not its model.

Sonnet ChatGPT Gemini Spark Deepseek Qwen
Memory What the agent can remember

By default an agent remembers your conversation up to a limit and forgets anything mentioned earlier. You can give it more memory like documents, past conversations or research papers and it will behave like a personal librarian.

The trade-off: the more it remembers, the more it costs to run and the easier it loses focus.

Tools How the agent acts

Without tools the agent can only chat. Tools can let it search the web, read files, write spreadsheets. Think of tools as the agent's hands - they give it the ability to do things.

Autonomy How much it acts on its own

Some agents ask for permission at every step while others just run until the job is done. The right level of autonomy depends on what's at stake. You wouldn't let an agent send emails to 10,000 students without a human checking the draft first.

Iteration How it works through a problem

It doesn't just answer once and stop. It tries something, checks if it worked, and decides what to do next, looping until the goal is met. This is what separates an agent from a chatbot. A chatbot responds. An agent persists.

↯ hover the rows ↯ foundations

A non-agentic workflow is a recipe.
An agent is a cook.

Non-agentic / Automation

predictable
  • One-pass execution
  • Predefined, fixed steps
  • No revisions, no iteration
  • "If X → do Y" rules
  • think: zero/one-shot workflows like summarising PDFs, rewriting emails, generating images
vs

Agentic

adaptive
  • ReasoningReasons about what to do next
  • MemoryRemembers past interactions
  • Access to toolsUses tools to act on the world
  • IterationIterates until goal is met
  • think: a digital employee
⟳ drag to turn ⟳ from prompt to autopilot

Autonomy is a dial, not a switch.

ZERO ORONE-SHOT HUMAN INTHE LOOP AGENTIC FULLYAUTONOMOUS

Use case
    Industries
      ↯ answer to navigate ↯ making the call

      Decision Guide:
      Agent or Automation?

      Question 1 of 2

      Can you map out every step the system should take, in advance?

      i.e., could you draw the whole flowchart before writing a line of code?

      Question 2 of 2

      Does the work mainly involve interpreting messy human content?

      emails, documents, chat transcripts, free-form text, images.

      Question 2 of 2

      Would quality improve if the system critiques its own draft and tries again?

      a self-review loop - draft, check, revise - before returning an answer.

      Recommendation

      Traditional automation.

      Predictable steps on structured data. You don't need an LLM - a script or if/then rules will be faster, cheaper, and more reliable.

      use cases Sending a confirmation email when a form is submitted Generating a weekly sales report on a schedule Moving data from one system to another
      Recommendation

      Use an AI step inside a workflow.

      You know the steps, but one of them needs to make sense of messy human content. Add one AI call into your pipeline - you don't need a full agent.

      use cases Reading an incoming email and deciding which team to send it to Pulling key details out of a contract or PDF Turning a long email thread into a short summary ticket
      Recommendation

      Build a tool-using agent.

      The path is open-ended - the system has to figure out each step as it goes, take actions, and decide what to do with the results.

      use cases A support assistant that looks up orders and issues refunds A research helper that searches the web and pulls together an answer A monitoring bot that spots problems and takes steps to fix them
      Recommendation

      Build a reflective agent.

      Open-ended path and a second pass makes the output better. The agent drafts, checks its own work, then revises before handing back the answer.

      use cases A coding assistant that writes code, runs it, and fixes the errors A research tool that writes a report, then checks it for gaps A writing assistant that drafts, then scores and rewrites its own output
      designing agents

      Agent Design Canvas.

      Before you build, fill in the text for this design canvas to form a complete brief for your agent. Flip each card to see how the six elements work together in this example of a billing support assistant that reads customer emails and decides what to do next.

      flip each card to fill the canvas
      01Role
      Who - or what - is this agent?
      flip to reveal ↓
      Name the job title, domain, and authority level. Be precise - the model fills every gap you leave.
      e.g. Billing support specialist. Handles inbound subscription billing disputes and account queries. Acts as first-line customer support.
      Skip it → Scope expands unpredictably. The model defaults to agreeable and overpromising.
      02Task
      What is it trying to accomplish?
      flip to reveal ↓
      State one output. Name the input, the transformation required, and what a correct result looks like. One task per agent.
      e.g. Read an inbound billing email. Classify the issue, set urgency 1-3, draft a reply, and flag if a human needs to step in.
      Skip it → The agent self-defines its goal. It optimises for something - just not necessarily what you intended.
      03Context
      What does it know going in?
      flip to reveal ↓
      List every data source by name: what's in the prompt, what it retrieves via tools, and what state it carries between turns.
      e.g. The customer's email, their account history from Salesforce, the refund policy doc, and their last 3 support tickets.
      Skip it → Data gaps get filled with confident-sounding invention. The agent won't flag what it doesn't know.
      04Tools
      What can it connect to?
      flip to reveal ↓
      Name every system it can look things up in or take action in. State what it can look up versus what it can change.
      e.g. Look up records in Salesforce, search the help centre, create tickets in Zendesk. Cannot edit billing records directly.
      Skip it → It only knows what it was trained on. No live systems, no actions.
      05Guardrails
      What must it never do?
      flip to reveal ↓
      Write explicit rules: what it must never do, what requires human approval before acting, and what triggers escalation.
      e.g. Never promise a refund without checking account status. Escalate if refund > $200, legal is mentioned, or 3+ open tickets.
      Skip it → The agent acts on best judgement. One edge case triggers an irreversible action.
      06Output
      What should the result look like?
      flip to reveal ↓
      Describe exactly what it should hand back - the format, what each part contains, and what a correct result looks like.
      e.g. Issue type, urgency (1-3), draft reply, escalate (yes/no), and reason if escalating. No free-form summaries.
      Skip it → The agent decides the format itself. If it varies, anything reading it downstream breaks.
      canvas complete - ready to build
      how agents are structured

      Agentic Design Patterns.

      / 04

        Best for

        Examples
        ways to organise agents

        Multi-agent Systems.

        / 04

        When to reach for it

        ↯ watch it refine ↯ measuring agent quality

        Evaluations: Measuring Agent Accuracy & Effectiveness.

        inputs agent outputs score 80% 100% 04 / 05 passing · fix the failing case 05 / 05 passing · regression fixed
        catch → fix → re-run

        How to run Evals

        1. 01
          Pick 10 real examples. Actual requests from real users - the inputs your agent has to handle day to day.
        2. 02
          Write down the expected result. For each one, note the answer or action you'd want back. That's your ground truth.
        3. 03
          Run and score the output. Compare each result to your ground truth - turn it into a percentage using a metric below.
        4. 04
          Refine and re-run. Tweak the prompt, swap the tool, change the flow - re-run the same examples and watch the score move.
        5. 05
          Grow the sample set over time. Every new failure becomes the next example. The broader the set, the more real the score.

        What to measure

        • Task success passes / total × 100%

          e.g. support agent resolves a refund request end-to-end

        • Trajectory sensible paths / total × 100%

          e.g. research agent actually reads the papers, not just titles

        • Tool accuracy correct tool calls / total × 100%

          e.g. sales agent queries the right CRM field with the right filters

        • Cost & time avg. $ & seconds per run

          e.g. coding assistant holds $0.12 / 8s avg, not drifting to $0.40 / 30s

        • Human-intervention rescued runs / total × 100%

          e.g. how often a human rep has to take over a support conversation

        ↯ sort · filter ↯ the agent stack

        Tools for building agents.

        Use
        Build
        Lift
        License
        Tool What it is Pick it for… No-code Lift
        Claude / CoworkGeneral Strong reasoning assistant plus an agentic desktop that works across your files and apps. Knowledge workers who need deeper reasoning, polished writing, or a desktop agent across files and apps. Yes Low–Med
        Gemini + WorkspaceGeneral AI built into Gmail, Docs, Meet, plus Gemini app and NotebookLM through Workspace plans. Orgs already in Google - meeting notes, docs, slides, and classroom prep in one plan. Yes Low
        Perplexity ComputerGeneral Cloud-hosted agentic task runner that coordinates 19 AI models to browse the web, control apps, fill forms, and complete multi-step workflows from natural language. Business users who want to delegate browser-based research and multi-step workflows in plain English - no setup, runs entirely in the cloud. Yes Low
        NotebookLMGeneral Source-grounded research and learning tool that works on your uploaded materials. Grounded learning and teaching prep - synthesise, query, and audio-overview your own sources. Yes Low
        Zapier AgentsNo-code No-code agents connected to thousands of apps, designed to act on business data and triggers. Quick no-code automations connecting business apps - the fastest path from idea to running agent. Yes Low
        GumloopNo-code Visual, AI-native agent builder where you drag-and-drop nodes (130+ integrations, MCP support) to create reasoning agents and multi-step automations, with a meta-agent that scaffolds flows from plain-language prompts. Teams wanting AI-native automation over simple triggers - agents that reason mid-flow, not just chain apps; step up from Zapier when logic and branching matter. Yes Low–Med
        n8nNo-code Workflow automation for technical teams with AI agents, traceable reasoning, visual builder, and self-hosting. Technical teams bridging workflow automation and real agent systems, with self-hosting options. Yes +code Medium
        Relevance AINo-code Low/no-code platform for building AI agents and multi-agent teams, especially around GTM and operations. Sales, onboarding, and GTM teams building AI agent workflows without writing code. Yes Medium
        MindStudioNo-code No-code visual builder for designing, deploying, and managing AI agents and apps across 200+ models and 1,000+ integrations - publishable as web apps, APIs, browser extensions, or scheduled automations. Teams and agencies building user-facing AI products without engineers - step above prompt wrappers, with 150K+ deployed agents and an approachable visual interface. Yes Low
        OpenAI CodexDeveloper Autonomous coding agent - available as web UI, desktop app, open-source CLI, and IDE extensions - that writes features, fixes bugs, runs tests, and orchestrates parallel sub-agents in sandboxed cloud environments. Engineering teams wanting an agent that writes, tests, and ships code autonomously - hands-off software engineer, accessible via CLI, API, or ChatGPT web UI. Partial Medium
        LangChainDeveloper Developer framework and platform (LangGraph + LangSmith) for building, testing, observing, and deploying custom agents. Developers building bespoke agent systems with full control over evals, memory, and observability. No High
        SierraEnterprise Enterprise customer-experience AI platform with agent studio, SDK, live assist, voice, and trust tooling. Enterprise CX teams deploying branded, voice-ready customer agents at scale. Platform-led Med–High
        OpenHandsOpen-source Open platform for cloud coding agents. Formerly OpenDevin; now stewarded by All Hands AI. Software engineering agents and sandboxed dev workflows - the leading open-source coding agent. No High
        OpenClawOpen-source Open-source self-hosted personal AI assistant with tool use, plugins, browser/canvas/actions, and companion apps. Power users building a personal agent layer with plugins, browser actions, and self-hosting. Partial Med–High
        Hermes AgentOpen-source Self-improving open-source agent by Nous Research with persistent memory, reusable skills, and a learning loop. Long-running autonomous agents that need cross-session memory and self-improving skills. Partial Med–High
        No tools match your filters.
        ↯ hands on ↯ the tool we're using

        Building on n8n.

        where it sits

        you are here Visual orchestrators n8n · Gumloop · MindStudio
        Agent frameworks LangGraph · LangChain · Semantic Kernel
        Model SDKs Anthropic SDK · OpenAI SDK · Vertex AI
        The models Claude · GPT · Gemini · Llama

        why n8n

        Open source Self-hostable Free at scale Exposes the logic Skills transfer

        Patterns here map to Gumloop, MindStudio, and LangGraph.

        what we're building

        research assistant
        • Take a topic as input
        • Search the web for recent sources
        • Summarise the findings with an LLM
        • Email you a brief

        Four nodes. One LLM call. One hour.

        before we start

        • n8n account created (pre-work)
        • API keys ready (provided)
        • Acronyms handout in your folder

        If something breaks, that's the lesson. Don't skip it.

        ↯ click the tabs ↯ hands-on build

        Build your first AI agent. In n8n.

        Mode A Research Agent

        Subject: research: transformer architecture

        Searches the web via Tavily, pulls the top 5 sources, synthesises them into a tight 200-word brief with 3 key takeaways. Emails it back to you.

        Mode B Lesson Agent

        Subject: teach me: vector databases

        Returns a structured 5-minute lesson - concept explained in plain English, one worked example, and a single practice exercise. Emails it back.

        One workflow. Two modes. Routing is just a subject-line check.

        Hostimap.gmail.com
        Port993
        Useryour@gmail.com
        Passwordapp password ← not your login password
        MailboxINBOX
        ActionMark as read
        Poll every1 minute
        1. Go to myaccount.google.com
        2. Security → 2-Step Verification must be on
        3. Search "App passwords" → generate one
        4. Name it n8n - copy the 16-char code
        Also enable IMAP in Gmail: Settings → See All Settings → Forwarding & POP/IMAP → Enable IMAP
        📧
        Email
        Trigger
        IF
        node
        True
        🔍
        HTTP
        Request
        Tavily
        🤖
        AI Agent
        research
        False
        🤖
        AI Agent
        lesson
        📤
        Send
        Email

        IF condition: {{ $json.subject.toLowerCase().startsWith('research') }} → True branch runs web search; False runs lesson generator

        01
        Email Trigger (IMAP)

        Add node → connect Gmail credentials (app password). Set poll: every 1 minute. This is your workflow entry point.

        02
        IF node

        Value 1: {{ $json.subject.toLowerCase() }}
        Operation: Contains · Value 2: research

        03
        HTTP Request - True branch

        URL: https://api.tavily.com/search · POST
        Body: {"query":"{{ $json.subject.replace('research:','').trim() }}","max_results":5}

        04a
        AI Agent - Research

        System prompt: "Synthesise these search results into a 200-word brief with 3 key takeaways. Be concise."

        04b
        AI Agent - Lesson (False branch)

        System prompt: "Create a 5-minute lesson on {{ topic }}: plain-English concept, one worked example, one practice exercise."

        05
        Send Email (Gmail)

        To: {{ $json.from.emailAddress.address }}
        Subject: Re: {{ $json.subject }}
        Body: {{ $('AI Agent').item.json.text }}

        1. Click Test workflow (top right)
        2. n8n begins polling your inbox every minute
        3. Send the test email from another account
        4. Watch each node light up as it executes
        5. Click any node → inspect Input / Output data
        6. Check your inbox for the reply email
        If a node errors: click it → read the Error tab. The message is usually literal. Fix it there.
        Research mode
        Subject: research: transformer architecture
        Body: (leave empty - not used)
        Lesson mode
        Subject: teach me: vector databases
        Body: (leave empty - not used)
        Edge case - should do nothing
        Subject: hello there
        Verify IF node takes False branch correctly
        💬
        Slack output

        Swap Send Email for a Slack node. Post the AI response to a channel. Entire team sees it.

        📝
        Log to Notion

        After Send Email, add a Notion node. Create a database row: topic, summary, timestamp.

        🕐
        Scheduled trigger

        Swap email trigger for Schedule node. Run every morning at 7am with a hardcoded topic.

        🧠
        Add memory

        Connect a Simple Memory node to the AI Agent. It now remembers your previous queries.

        Every extension is one more node. That's the point.

        live workshop session

        Real client. Real problem.
        Forty minutes. One good agent.

        Your rules of engagement

        • Get into groups.
        • Pick one problem worth solving for the business.
        • Plan your agent on the canvas before you touch n8n.
        • Build a working proof of concept in n8n.
        • You have 40 minutes for the above!
        • Present to the client. Five minutes per group.

        What you'll need to define for your agent

        • Agent goal (Use the Agent Design Canvas)
        • Inputs & outputs
        • Data sources (structured vs unstructured)
        • Human-in-the-loop checkpoints (if any)
        • Evaluation method

        At the end each team shares: problem, design, one surprise.

        workshop

        Workshop: Agent Design Canvas.

        Fill in each card to form the brief for the agent you want to build.

        0 / 6 fields filled
        01Role
        Who - or what - is this agent?
        02Task
        What is it trying to accomplish?
        03Context
        What does it know going in?
        04Tools
        What can it connect to?
        05Guardrails
        What must it never do?
        06Output
        What should the result look like?
        to close

        Tools change,
        knowledge compounds.

        The meta-skill is learning to evaluate, switch, and compose tools - not mastering any one. What you built today is the muscle for everything that comes next.

        thanks for coming ↯
        doodle mode on · press D or esc to exit