5-Day Programme

NYMG Agentic Academy

AI is changing how we work. This week, you'll understand what it is, how it works, how we use it, and what you're expected to learn while the systems are still evolving - in five short daily sessions.

1
LLMs
2
Landscape
3
Agents
4
Setup
5
Your Role

👋 Welcome to Learning Week

Over the next five days, you'll go from "what even is AI?" to understanding how the team uses it, where agents fit, and what good operator behaviour looks like.

Each day takes about 20 minutes: a short podcast, flashcards to test yourself, a quick quiz, and reflection questions for the team meeting.

No tech background needed. Expect a bumpy ride, not a polished fantasy. Just listen, learn, question what you see, and come with examples from your own work.

← Back to overview
1
Day 1 of 5

What Are LLMs and Why Should You Care?

In progress
Step 1 of 6 - Listen
🎙️

Listen First

Start with the podcast. Two hosts walk you through everything you need to know about LLMs - what they are, how they work, and why they sometimes get things wrong.

⏱ ~14 minutes

What You'll Learn

  • The ChatGPT moment
  • How LLMs are trained
  • Prediction vs thinking
  • LLM ≠ computer
  • Tokens & costs
  • Context window
  • Knowledge cutoff
  • Hallucination
Step 2 of 6 - Study
🎨

Study the Visual

See how LLM prediction actually works - from training data to next-token output.

Step 3 of 6 - Flashcards
Flashcards - Day 1
1 / 13
Term
LLM
tap to reveal →
Large Language Model. A mathematical pattern-recognition system trained on enormous amounts of text. It predicts the most likely next word - it doesn't think or reason.
💡 The world's most sophisticated autocomplete. Like your phone keyboard, but trained on most of the internet.
🔍 Deeper Dive
Question
Why do LLMs hallucinate?
tap to reveal →
Because they're optimised for fluency, not truth. They're rewarded for producing text that sounds good. They have no fact-checker and no confidence meter - they can't evaluate their own certainty.
Step 4 of 6 - Quiz
🧠 Quick Quiz - Day 1 Question 1/3
What does an LLM do when generating text?
Why do LLMs hallucinate?
What is a context window?
Step 5 of 6 - Review
📌 Key Takeaways
  • An LLM is a prediction engine, not a thinking machine
  • It's trained on text data - that data IS its worldview
  • It's probabilistic (not a computer) - optimised for fluency, not truth
  • Everything runs on tokens, with a hard context window limit
  • On its own: no internet, no memory, no action

The ChatGPT Moment

In November 2022, OpenAI released ChatGPT. The technology wasn't new - Google invented it in 2017 - but the simple chat interface made it accessible to everyone. Like the web browser did for the internet.

How LLMs Are Built

An LLM is trained on enormous amounts of text - books, websites, papers, code. This training data IS its entire worldview. It has read about the world but never experienced it.

Predicts, Doesn't Think

An LLM predicts the most likely next word based on patterns. It's the world's most sophisticated autocomplete. It does not reason or understand.

Not a Computer

Computers are deterministic - 2+2 always = 4. LLMs are probabilistic - they deal in likelihood, not certainty. That's why they can't reliably count the R's in "strawberry" (they see tokens, not individual letters).

Tokens & Context Window

A token ≈ ¾ of a word. Everything costs tokens. The context window is the AI's desk - everything must fit. Some models handle up to 1 million tokens (~1,500 pages), but instructions, conversation, and answers all share that space.

Knowledge Cutoff

LLMs are frozen in time after training. They don't learn from new events. Ask about last week's news and they'll either say "I don't know" or confidently make something up.

Hallucination

When AI generates confident but completely made-up information. It's optimised for fluency, not truth. It has no fact-checker and no humility - it can't say "I'm not sure."

What LLMs Cannot Do (On Their Own)

  1. Browse the internet - frozen at training cutoff
  2. Remember past conversations - every chat starts from zero
  3. Take action - can write an email but can't send it
💡 Key takeaway: An LLM is a prediction engine, not a thinking machine. Always verify its output.
  • Have you used ChatGPT or another AI tool? What was your experience?
  • Can you think of a time you trusted information that turned out to be wrong? How does that relate to hallucination?
  • What surprised you most about how LLMs actually work?
  • What's one thing you'd want to verify before trusting an AI answer?
LLM (Large Language Model)
A mathematical pattern-recognition system trained on enormous amounts of text. It predicts the most likely next word - it doesn't think or reason.
Training Data
The text an LLM learned from (books, websites, papers, code). This data IS the model's entire worldview.
Token
The basic unit an LLM works with, roughly ¾ of a word. Everything in AI has a token cost.
Context Window
The maximum tokens an AI can handle at once. Like a desk - if it overflows, the oldest information falls off.
Context
Everything the AI can "see" right now. If it's not in the context, it doesn't exist for the AI.
Knowledge Cutoff
The date when an LLM's training data stops. After this, the model is frozen in time.
Hallucination
When AI generates confident but entirely made-up information. It has no fact-checker and no humility.
Probabilistic vs Deterministic
Computers are deterministic (same input = same output). LLMs are probabilistic (deal in likelihood, not certainty).
Prompt
The text you send to an AI. Better input = better output.
Transformer
The architecture behind modern LLMs. Invented by Google in 2017, made accessible by OpenAI in 2022.
Step 6 of 6 - Complete
🎉

Day 1 Complete!

You've mastered the fundamentals. Before moving on - see how fast this all happened:

🎬 Bonus: Ryan Serhant on AI

You know him from Owning Manhattan - hear his take on AI and trust.

Tap play to load · Video may take a moment over the network

← Back to overview
2
Day 2 of 5

The Landscape: Who's Who and What to Use

AI is not one thing from one company. Different models for different jobs. Smart companies match the tool to the task.

🔒 Locked
🔒

Day 2 is locked

Complete Day 1 flashcards to unlock.

Finish Day 1 first →
Step 1 of 6 - Listen
🎙️

Listen First

Today we zoom out from the technology to the landscape. Who builds these AI models, why they're all different, and how NYMG navigates the choices.

⏱ ~10 minutes

What You'll Learn

  • The major AI companies
  • What "open source" means
  • Why models feel different
  • Model versions & upgrades
  • What model "size" means
  • Thinking / Doing / Checking
  • NYMG's AI journey
  • Passive vs Active AI
Step 2 of 6 - Study
🎨

Study the Visual

The wider AI landscape at a glance - and the simple setup NYMG uses in practice.

Step 3 of 6 - Flashcards
Flashcards - Day 2
1 / 13
Term
Provider
tap to reveal →
A company that builds and offers AI models. The major providers include OpenAI, Anthropic, Google, Meta, and xAI. Each has a different philosophy and approach.
💡 Like car manufacturers - BMW, Toyota, Tesla all make cars, but with very different philosophies.
🔍 Deeper Dive
Question
Why does the same question give different answers in ChatGPT vs Claude?
tap to reveal →
Because each model was trained differently and optimised for different goals. OpenAI is designed to make you feel good - warm and conversational. Anthropic is designed to make you think - more precise and careful. Neither is wrong.
Step 4 of 6 - Quiz
🧠 Quick Quiz - Day 2 Question 1/3
Why might you get different answers from ChatGPT vs Claude on the same question?
Thinking models (like claude-opus-4-6) are best suited for which type of task?
What is the difference between passive AI and active AI?
Step 5 of 6 - Review
📌 Key Takeaways
  • AI is not one company - it's a global ecosystem of providers
  • Different models feel different because they're optimised for different goals
  • Models come in tiers: thinking (expensive), doing (workhorse), checking (cheap)
  • NYMG uses multiple tools - Claude for heavy lifting, ChatGPT for helpdesk
  • Everything so far is passive AI - you ask, it answers

The Major Players

CompanyModelKnown For
OpenAIGPT / ChatGPTStarted the revolution. Biggest consumer AI.
AnthropicClaudeSafety-focused. Massive context window. NYMG's primary AI.
GoogleGeminiInvented the technology (2017). Integrated with Workspace.
MetaLlamaOpen source - free for anyone to download and use.
xAIGrokElon Musk's company. Growing quickly.
DeepSeekDeepSeekChinese. Open source. Competitive quality. Free.

Why Models Feel Different

Same question, three different answers. OpenAI is designed to make you feel good (warm, conversational). Anthropic is designed to make you think (precise, careful). Each model has its own personality based on training.

Models Change

AI models get version updates (claude-sonnet-4-6 replacing earlier Sonnet generations). Behaviour, formatting, and tone can shift. Companies track versions and test after updates. Benchmarks compare models objectively, but don't tell the full story.

Model Tiers - Thinking / Doing / Checking

TierExampleUse ForCost
Thinkingclaude-opus-4-6Complex reasoning, strategy$$$
Doingclaude-sonnet-4-6Everyday tasks (90% of work)$$
Checkingclaude-haiku-4-5Quick grammar, formatting$

Same task can cost 10× more on a thinking model vs a checking model.

How NYMG Uses AI

  1. Started on helpdesk → wasn't reliable
  2. Moved into technical operations and AI-assisted coding workflows
  3. A jump to newer Claude generations was the turning point for coding productivity
  4. Claude Code let Charlotte do tech tasks without coding knowledge
  5. Dev team shifted: writing code → reviewing AI-written code
  6. Bart uses ChatGPT for helpdesk; Charlotte & Daniëlle use Claude Code

Passive AI vs Active AI

Passive: You go to it. Ask a question, get an answer. (ChatGPT, Claude)

Active: It can take initiative - use tools, remember past conversations, work through steps independently.

💡 Key takeaway: AI is not one thing from one company. Different models for different jobs. Smart companies match the tool to the task.
  • Have you noticed different results when using different AI tools? What was different?
  • Which of the AI tools mentioned (ChatGPT, Claude) are you most curious to try?
  • Can you think of tasks in your daily work that could use a 'checking model' vs a 'thinking model'?
  • What does 'passive AI' vs 'active AI' mean to you based on what you heard?
Day 1 - What Are LLMs?
LLM (Large Language Model)
A mathematical pattern-recognition system trained on enormous amounts of text. It predicts the most likely next word - it doesn't think or reason.
Training Data
The text an LLM learned from (books, websites, papers, code). This data IS the model's entire worldview.
Token
The basic unit an LLM works with, roughly ¾ of a word. Everything in AI has a token cost.
Context Window
The maximum tokens an AI can handle at once. Like a desk - if it overflows, the oldest information falls off.
Context
Everything the AI can "see" right now. If it's not in the context, it doesn't exist for the AI.
Knowledge Cutoff
The date when an LLM's training data stops. After this, the model is frozen in time.
Hallucination
When AI generates confident but entirely made-up information. It has no fact-checker and no humility.
Probabilistic vs Deterministic
Computers are deterministic (same input = same output). LLMs are probabilistic (deal in likelihood, not certainty).
Prompt
The text you send to an AI. Better input = better output.
Transformer
The architecture behind modern LLMs. Invented by Google in 2017, made accessible by OpenAI in 2022.
Day 2 - The AI Landscape
Provider
A company that builds AI models (OpenAI, Anthropic, Google, Meta, xAI, DeepSeek).
Open Source
When a model is released openly, so anyone can download, use, and modify it.
Model Size
How many patterns/connections a model learned. Bigger = more capable but slower and more expensive.
Thinking / Doing / Checking Models
Three tiers: Thinking (claude-opus-4-6 - complex reasoning), Doing (claude-sonnet-4-6 - everyday tasks), Checking (claude-haiku-4-5 - quick, cheap quality control).
Model Version
AI models get updated regularly (e.g. claude-sonnet-4-6 replacing earlier Sonnet generations). Each version may change behaviour, formatting, and tone.
Benchmark
A standardised test for comparing AI models. Useful but doesn't tell the full story.
Model Routing
Sending each task to the right model tier. Simple tasks → cheap model, complex tasks → powerful model.
Passive AI
AI that waits for you to ask. You go to it, ask a question, get an answer.
Active AI
AI that can take initiative, use tools, remember conversations, and work through tasks independently.
Claude Code
Anthropic's tool that works directly with code. Different from Claude the chatbot - it can build, fix, and create.
Step 6 of 6 - Complete
🎉

Day 2 Complete!

You now know who builds AI models, why different tools exist for different jobs, and how to spot agent washing. Tomorrow we cross the line from passive to active AI.

← Back to overview
3
Day 3 of 5

From Chatbot to Agent

🔒 Locked
🔒

Day 3 is locked

Complete Day 2 flashcards to unlock.

Finish Day 2 first →
Step 1 of 6 - Listen
🎙️

Listen First

Today we cover the most important concept of the week: the difference between a chatbot and an agent. This is where passive AI becomes active AI - and where everything clicks.

⏱ ~12 minutes

What You'll Learn

  • Chatbot vs agent
  • Tools (the agent's hands)
  • Memory & autonomy
  • System prompts explained
  • Guardrails & safety
  • The agent loop
  • Workflow automation
  • Orchestrators
🤖

Bonus: What is an Agent?

A short explainer. The anatomy of an agent in 6 building blocks: model, memory, skills, tools, guardrails, identity.

⏱ ~4 minutes
Step 2 of 6 - Study
🎨

Study the Visual

The key shift: from a chatbot that answers, to an agent that acts.

Step 3 of 6 - Flashcards
Flashcards - Day 3
1 / 17
Term
Chatbot
tap to reveal →
An AI that can have conversations but can't take action. It responds to what you ask, but it can't go do things in the world.
💡 Like a receptionist who answers questions but can't leave the desk.
🔍 Deeper Dive
Question
What three things transform a chatbot into an agent?
tap to reveal →
Tools (capabilities to interact with the world), Memory (short-term context + long-term database), and Autonomy (ability to decide steps, guided by a pipeline).
Step 4 of 6 - Quiz
🧠 Quick Quiz - Day 3 Question 1/3
What three things transform a chatbot into an agent?
What is the purpose of guardrails in an AI agent?
What best describes an agent loop?
Step 5 of 6 - Review
📌 Key Takeaways
  • Agentic = able to take action. A chatbot talks, an agent acts. Watch out for agent washing.
  • Three pillars: tools (capabilities), memory (short-term + long-term), guided autonomy (freedom within a pipeline)
  • Tools are individual actions. Skills are structured instructions that orchestrate multiple tools for complex tasks.
  • Guardrails keep agents safe - rules + pipeline checkpoints. Joey's duplicate story proves why.
  • 2026 = the agentic shift. Every major company is building agent capabilities. NYMG is part of this.

Chatbot vs Agent

A chatbot is passive AI - you ask, it answers, but it can't take action. An agent is active AI that can use tools, remember past interactions, and work autonomously toward goals.

The Three Pillars of an Agent

PillarWhat It MeansExample
ToolsExternal capabilities to interact with the worldBrowse web, edit files, send messages
MemoryStores and recalls past conversationsRemembers your preferences from last week
AutonomyPlans its own steps without instructionFigures out how to complete a goal independently

System Prompts

Hidden instructions that define an AI's persona, rules, and boundaries. The user never sees them, but they shape everything. The same AI becomes a completely different tool with different system prompts - that's how Claude becomes Claude Code, or a customer support bot.

Guardrails

Predefined rules and limits that prevent an agent from taking dangerous actions. Without guardrails, autonomy is a liability. With guardrails, it's a superpower.

The Agent Loop

  1. Observe - take in the current situation
  2. Think - decide what to do next
  3. Act - use a tool or take a step
  4. Check - evaluate the result
  5. Repeat until the job is done

Workflow Automation & Orchestrators

When multiple specialist agents work together under an orchestrator, you get workflow automation - entire processes running from start to finish. At NYMG, Jørgen's mystery tool from Day 2 is an agent, and the company is moving toward specialised agents coordinated by an orchestrator.

💡 Key takeaway: A chatbot talks. An agent acts. Tools + Memory + Autonomy = Agent.
  • Think about your daily work: which tasks do you do repeatedly that follow the same steps each time?
  • If you could give an AI assistant three tools to help with your job, what would they be?
  • What's one task where you'd want to keep human approval (guardrails) and one where you'd trust the AI to just do it?
  • How does the chatbot vs agent distinction change how you think about AI in your work?
Day 1 - What Are LLMs?
LLM (Large Language Model)
A mathematical pattern-recognition system trained on enormous amounts of text. It predicts the most likely next word - it doesn't think or reason.
Training Data
The text an LLM learned from (books, websites, papers, code). This data IS the model's entire worldview.
Token
The basic unit an LLM works with, roughly ¾ of a word. Everything in AI has a token cost.
Context Window
The maximum tokens an AI can handle at once. Like a desk - if it overflows, the oldest information falls off.
Context
Everything the AI can "see" right now. If it's not in the context, it doesn't exist for the AI.
Knowledge Cutoff
The date when an LLM's training data stops. After this, the model is frozen in time.
Hallucination
When AI generates confident but entirely made-up information. It has no fact-checker and no humility.
Probabilistic vs Deterministic
Computers are deterministic (same input = same output). LLMs are probabilistic (deal in likelihood, not certainty).
Prompt
The text you send to an AI. Better input = better output.
Transformer
The architecture behind modern LLMs. Invented by Google in 2017, made accessible by OpenAI in 2022.
Day 2 - The AI Landscape
Provider
A company that builds AI models (OpenAI, Anthropic, Google, Meta, xAI, DeepSeek).
Open Source
When a model is released openly, so anyone can download, use, and modify it.
Model Size
How many patterns/connections a model learned. Bigger = more capable but slower and more expensive.
Thinking / Doing / Checking Models
Three tiers: Thinking (claude-opus-4-6 - complex reasoning), Doing (claude-sonnet-4-6 - everyday tasks), Checking (claude-haiku-4-5 - quick, cheap quality control).
Model Version
AI models get updated regularly (e.g. claude-sonnet-4-6 replacing earlier Sonnet generations). Each version may change behaviour, formatting, and tone.
Benchmark
A standardised test for comparing AI models. Useful but doesn't tell the full story.
Model Routing
Sending each task to the right model tier. Simple tasks → cheap model, complex tasks → powerful model.
Passive AI
AI that waits for you to ask. You go to it, ask a question, get an answer.
Active AI
AI that can take initiative, use tools, remember conversations, and work through tasks independently.
Claude Code
Anthropic's tool that works directly with code. Different from Claude the chatbot - it can build, fix, and create.
Day 3 - From Chatbot to Agent
Chatbot
An AI that can converse but can't take action. Like a receptionist who can't leave the desk.
Agent
An AI that can plan and use tools to complete tasks. It doesn't just answer - it acts.
Tools
Individual capabilities given to an agent - browse web, read files, send messages. Each tool is a single action.
Skills
Structured instructions that orchestrate multiple tools for complex tasks. A tool is a single action; a skill coordinates many.
Short-Term Memory
The current conversation context, limited by the context window. Like your desk - visible but finite.
Long-Term Memory
Information stored in an external database, searchable across conversations. Like a filing cabinet.
Autonomy & Pipeline
Autonomy = deciding own steps. Pipeline = designed sequence of steps. Most real agents use guided autonomy within pipelines.
Guardrails
Rules + pipeline checkpoints preventing harmful actions. Autonomy without guardrails = real problems.
Agent Loop
Observe → think → act → check, repeating until the job is done.
Sub-Agent
A child agent spawned for a specific subtask. Like hiring a freelancer for one job.
Agent Washing
Labelling traditional automation as "agentic." Real agents reason and adapt; fake agents follow fixed scripts.
Computer Use
AI capability to see your screen and control mouse/keyboard. Introduced by Anthropic, being built by all major companies.
Anthropomorphism
Giving AI human-like traits (name, personality). Increases engagement but the AI isn't actually feeling anything.
The Agentic Shift
The 2026 industry-wide move from information-providing AI to task-executing agents.
Step 6 of 6 - Complete
🎉

Day 3 Complete!

You now understand the shift from chatbot to agent. Day 4 is now unlocked - time to look under the hood.

← Back to overview
4
Day 4 of 5

Setup

🔒 Locked
Step 1 of 6 - Listen
🎙️

Listen First

The most NYMG-specific episode. See how Atlas, Joey, and the other agents work behind the scenes, hear real war stories, and understand how everything connects.

⏱ ~19 minutes

What You'll Learn

  • Why OpenClaw (local, open-source)
  • SOUL file and more - how agent identity works
  • The glossary - 15 years of existing institutional knowledge, used by agents automatically
  • Meet Atlas - your primary day-to-day agent
  • How Atlas uses skills, pipelines, and approval checkpoints
  • How glossary knowledge is applied across markets
  • Memory, heartbeats, and cron jobs (at a practical level)
  • War stories - what went wrong and why
  • Anthropomorphism - agents have names, not feelings
Step 2 of 6 - Study
🎨

Study the Visual

Three visuals: how an agent is built, the overall system, and Atlas's content pipeline.

Step 3 of 6 - Flashcards
Flashcards - Day 4
1 / 16
Term
SOUL.md
tap to reveal →
The personality file that shapes an agent's identity. Joey's says "Real New Yorker energy." Atlas says "Named for the titan who carries the world." It defines who the agent IS - its tone, rules, and character.
💡 Like a character bible for a TV show - every writer reads it so the lead always sounds like themselves.
Term
Workspace Files
tap to reveal →
A collection of identity files that together form an agent's working instructions in OpenClaw. Not one text box - organised, readable files.
💡 Like a new employee's onboarding folder: who you are, who you serve, what tools you use - all in separate tabs.
Term
Pipeline
tap to reveal →
A designed sequence of steps built in LangGraph. Atlas's current production pipeline is: parse_request → fetch_content → assess_and_plan → translate → copy_edit → glossary_check → apply_edit → validate → replicate_sites → report.
💡 Like an airport conveyor belt - your bag passes through check-in, security, and baggage claim in a fixed order.
Term
Human-in-the-Loop
tap to reveal →
Human approval at key decision points. Jørgen reviews Joey's posts before they go live. Country managers review Atlas's translations. The AI does the work, humans approve it.
💡 Like a sous chef who preps everything - but the head chef tastes and nods before it leaves the kitchen.
Term
Sub-Agent
tap to reveal →
One-shot - created for one task only, no memory, no continuity. Atlas spawns one per language: it gets the task, glossary, and style guide - does the full pipeline, then disappears. You only ever talk to main Atlas.
💡 Like a temp contractor hired for one job - they arrive, do the work, and leave. No ongoing relationship.
Term
Heartbeat
tap to reveal →
A scheduled wake-up where the agent proactively runs a checklist - check inbox, run health checks, monitor token usage. No one asks it to. It works while you sleep.
💡 Like a night watchman doing rounds - checking every door on schedule, not waiting to be called.
Term
Cron Job
tap to reveal →
A scheduled task at a specific time. "Post this reel at 9 AM Saturday." Joey's crons don't survive restarts - there's a manifest file to recreate them.
💡 Like a calendar alarm - set it once, it fires at exactly the right moment without you being awake.
Term
Compaction
tap to reveal →
When a conversation gets too long for the context window, the system summarises older parts - like meeting minutes instead of a full transcript. The agent still knows what happened, in condensed form.
💡 Like a book summary - you didn't re-read every page, but you know the plot well enough to keep going.
Term
pgmemory
tap to reveal →
The team's long-term memory system. A PostgreSQL database that auto-captures important information and auto-recalls relevant memories in new conversations. The glossary is now query-based in pgmemory with 7,283 entries.
💡 Like a colleague who's been here 10 years - they just know things without being told, because they were there.
Term
Atlas Workflow
tap to reveal →
The practical path country managers see: task comes in, Atlas applies the change, quality checks run, and you review the result before anything is final.
💡 Like tracked changes in a Word doc - Atlas edits, you accept or reject, nothing goes live without your nod.
Term
Fallback Chain
tap to reveal →
A backup plan when the primary AI model fails. The system automatically switches to approved alternatives so work can continue without interruption.
💡 Like a generator kicking in when the power cuts - you barely notice, the lights stay on.
Term
Glossary Governance
tap to reveal →
The glossary captures years of institutional knowledge. It now lives in a pgmemory table, so the agent queries the right terms instead of loading giant files into context.
💡 Like a company style guide that's searchable online - you look up what you need, not carry the whole book.
Quiz
How is the "system prompt" set up in OpenClaw?
tap to reveal →
Through workspace files: SOUL.md (personality), AGENTS.md (rules), USER.md (communication style), MEMORY.md (context), and TOOLS.md (tool guidance). Together they form the working prompt - organised into readable files, not one text box.
Quiz
Why does the heartbeat matter for agent reliability?
tap to reveal →
It lets agents work proactively - checking health, monitoring tasks, catching problems before anyone notices. Without heartbeats, agents only respond when spoken to. With heartbeats, they watch over things while you sleep.
Quiz
What should you do if Atlas gives a wrong translation?
tap to reveal →
Flag it and correct it. Your expertise as a country manager protects quality in your market. The glossary captures 15 years of NYMG institutional knowledge, and Atlas relies on that accuracy every day.
Quiz
What's the difference between a guardrail-as-instruction and a guardrail-as-code?
tap to reveal →
An instruction tells the AI "don't do this" - but the AI can ignore it. Code physically blocks the action - like the duplicate checker that stops uploads before they happen. Code guardrails are stronger than instruction guardrails.
Step 4 of 6 - Quiz
🧠 Quick Quiz - Day 4 Question 1/3

How is the "system prompt" configured in OpenClaw?

What happens when Atlas needs to translate a post into 17 languages?

What's the key difference between a guardrail-as-instruction and a guardrail-as-code?

Step 5 of 6 - Review

📌 Key Takeaways

  • Local = control. OpenClaw runs on Charlotte's Mac Mini. Your data stays in the building.
  • System prompt = files. SOUL.md (personality), AGENTS.md (rules), USER.md (comms), MEMORY.md (context), TOOLS.md (tool guidance). Not a single text box.
  • Atlas is your main agent. It handles content updates, translations, and quality checks for your market. Joey handles social media in a separate workflow.
  • Skills auto-load. The right skill activates when the task matches - the agent doesn't need to be told which skill to use.
  • Memory persists. pgmemory auto-captures and auto-recalls. Compaction summarises long conversations.
  • Heartbeats = proactive. Agents check in without being asked. They work while you sleep.
  • Guardrails: instruction vs code. Instructions can be ignored. Code gates physically block bad actions. We're upgrading from instructions to code.
  • Things go wrong. Duplicate posts, ignored stops, crashes. That's normal. That's why agents need humans.

OpenClaw - A local, open-source platform running on Charlotte's Mac Mini as a separate user. Multi-agent by design. Data stays on our hardware.

System Prompt in Practice - Not a text box. Workspace files: SOUL.md defines personality, AGENTS.md sets rules, USER.md teaches communication style, MEMORY.md provides bootstrap context, and TOOLS.md explains tool usage.

The Agents - Atlas 🗺️ is the one country managers work with most (content ops, 17 languages, sub-agents per language). Joey 🗽 runs social workflows with Jørgen. Other technical agents exist in the background, but your day-to-day interaction is mainly Atlas.

Skills - Structured instructions that auto-load when tasks match. Atlas has 18 skills. Shared library means consistent processes across agents.

Memory - Short-term = current conversation (context window). Compaction = summarising when it gets long. Long-term = pgmemory (PostgreSQL, auto-capture, auto-recall).

Heartbeat - Proactive scheduled check-ins. Health checks, inbox monitoring. Joey: weekends only. Powered by cron jobs (scheduled tasks).

Pipelines (LangGraph) - Designed step sequences with checkpoints. Joey's posting pipeline. Atlas's translation pipeline. Hard gates physically block bad actions.

Fallback Chains - Automatic model switching when the primary AI provider is down. Ensures agents keep working.

War Stories - Duplicate reel (health monitor killed Joey mid-upload). Ignored STOP command (instruction vs code guardrail). These examples show why process and human review matter.

Glossary - Years of institutional knowledge, now stored in a pgmemory glossary table with 7,283 entries. Agents query what they need instead of loading giant per-language dumps.

  1. Which agent will you be working with most directly? What does it do?
  2. What's the difference between how Joey handles social media and how you'd do it manually?
  3. Why does it matter that NYMG's agents run locally instead of in the cloud?
  4. If Atlas makes a mistake in your language, what should you do?
SOUL.md
The personality file - defines who the agent IS. Tone, rules, character.
Workspace Files
SOUL.md + AGENTS.md + USER.md + MEMORY.md = the complete "system prompt" in OpenClaw.
Pipeline
A designed sequence of steps in LangGraph. Each step has checkpoints and can resume on failure.
Human-in-the-Loop
Human approval at key decision points before an agent proceeds.
Sub-Agent
A temporary child agent spawned for one task. Gets the right glossary, does the job, disappears.
Heartbeat
Proactive scheduled check-in. The agent works while you sleep.
Cron Job
A scheduled task at a specific time. Don't survive restarts - recreated from manifests.
Compaction
Summarising long conversations to free up context window space.
pgmemory
PostgreSQL-based long-term memory. Auto-captures, auto-recalls. Not perfect but better than starting from zero.
Atlas Workflow
The practical sequence country managers see: receive task, apply change, run checks, review output, then confirm.
Fallback Chain
Automatic backup models when the primary AI provider is unavailable. Keeps agents working.
Practice: Meet Atlas

🗺️ Meet Atlas - Your Content Agent

Atlas is the agent you'll be working with most. It lives in your Slack channel and handles content updates, translations, and quality checks for your market.

Your Slack channel: #atlas-[your-country-code]
e.g., #atlas-nl for Dutch, #atlas-fi for Finnish, #atlas-de for German

Your First Task

🚧 Your first guided task will be provided in the meeting. Charlotte will walk you through it step by step.

⚠️ What to Do When Things Go Wrong

  • Wrong translation? Tell Atlas directly: "That's incorrect. It should be [correct term]."
  • Atlas isn't responding? Wait 30 seconds, then try again. If still nothing, message Charlotte.
  • Atlas did something you didn't ask? Tell it to stop and revert. Then flag it to Charlotte.
  • Not sure if it's right? Ask Atlas to show you what it changed before approving.
Step 6 of 6 - Complete
🎉

Day 4 Complete!

You now know how the agents work under the hood. Day 5 is now unlocked - the practical reality of working with them.

🎁 Bonus: Atlas Content Pipeline

A detailed look at every step Atlas follows when updating your pages.

← Back to overview
5
Day 5 of 5

Along for the Ride

In progress
Step 1 of 6 — Listen
🎙️

Final lesson

This final day is the operator's guide: how to brief agents clearly, how to verify their work, what they can see, and how your corrections make the system better over time.

⏱ ~8 minutes

What You'll Learn

  • Be specific, not clever — avoid assumption and ambiguity
  • Trust, but verify — always check the output
  • Your feedback trains the system permanently
  • What the agent can see — privacy and boundaries
  • When things go wrong — practical troubleshooting
  • Philosophy: try it, watch carefully, improve every week
Step 2 of 6 — Study
Step 3 of 6 — Flashcards
Flashcards — Day 5
1 / 6
Principle
Be specific, not clever
tap to reveal →
Good agent instructions are concrete. If a detail matters, include it. If something must stay untouched, say that too.
💡 Ambiguity creates guesses. Specificity reduces mistakes.
Step 4 of 6 — Quiz
🧠 Quick Quiz — Day 5 Question 1/5
What is the best way to brief an agent?
If an agent says "done," what should you assume?
Why is operator feedback so valuable?
What determines what an agent can see?
What is the best response when an agent produces a wrong translation?
Step 5 of 6 — Discuss
💬

Team Reflection

Bring these questions to your next team meeting. There are no right answers — just honest ones.

  • What surprised you most across all 5 days?
  • What's one thing you'll do differently starting next week?
  • Where do you see the biggest opportunity for agents in your specific role?
  • What's your biggest remaining concern or question?
Step 6 of 6 — Complete
🎓

You've completed the Agentic Academy!

Five days of learning, from "what is AI?" to working alongside agents every day. The goal now: informed use, better reviews, and clearer workflows.

📚 Glossary

Day 1 - What Are LLMs?
LLM (Large Language Model)
A mathematical pattern-recognition system trained on enormous amounts of text. It predicts the most likely next word - it doesn't think or reason.
Training Data
The text an LLM learned from (books, websites, papers, code). This data IS the model's entire worldview.
Token
The basic unit an LLM works with, roughly ¾ of a word. Everything in AI has a token cost.
Context Window
The maximum tokens an AI can handle at once. Like a desk - if it overflows, the oldest information falls off.
Context
Everything the AI can "see" right now. If it's not in the context, it doesn't exist for the AI.
Knowledge Cutoff
The date when an LLM's training data stops. After this, the model is frozen in time.
Hallucination
When AI generates confident but entirely made-up information. It has no fact-checker and no humility.
Probabilistic vs Deterministic
Computers are deterministic (same input = same output). LLMs are probabilistic (deal in likelihood, not certainty).
Prompt
The text you send to an AI. Better input = better output.
Transformer
The architecture behind modern LLMs. Invented by Google in 2017, made accessible by OpenAI in 2022.
Day 2 - The AI Landscape
Provider
A company that builds AI models (OpenAI, Anthropic, Google, Meta, xAI, DeepSeek).
Open Source
When a model is released openly, so anyone can download, use, and modify it.
Model Size
How many patterns/connections a model learned. Bigger = more capable but slower and more expensive.
Thinking / Doing / Checking Models
Three tiers: Thinking (claude-opus-4-6 - complex reasoning), Doing (claude-sonnet-4-6 - everyday tasks), Checking (claude-haiku-4-5 - quick, cheap quality control).
Model Version
AI models get updated regularly (e.g. claude-sonnet-4-6 replacing earlier Sonnet generations). Each version may change behaviour, formatting, and tone.
Benchmark
A standardised test for comparing AI models. Useful but doesn't tell the full story.
Model Routing
Sending each task to the right model tier. Simple tasks → cheap model, complex tasks → powerful model.
Passive AI
AI that waits for you to ask. You go to it, ask a question, get an answer.
Active AI
AI that can take initiative, use tools, remember conversations, and work through tasks independently.
Claude Code
Anthropic's tool that works directly with code. Different from Claude the chatbot - it can build, fix, and create.
Day 3 - From Chatbot to Agent
Chatbot
An AI that can have conversations but can't take action. Like a knowledgeable receptionist who answers questions but can't leave the desk.
Agent
An AI that can plan and use tools to complete tasks independently. Like a personal assistant who books your flights and hotels proactively.
Tools
External capabilities given to an agent to interact with the world (browse web, edit files, send messages). Like giving someone tools to work with, not just a manual.
Memory
How an agent stores and recalls past conversations to inform future decisions. Like a colleague who remembers your preferences from previous projects.
Autonomy
The ability to decide its own steps without constant human instruction. Like a delivery driver choosing the best route without turn-by-turn directions.
System Prompt
Hidden instructions that define an AI's persona, rules, and boundaries. Like a detailed job description that shapes how an employee behaves.
Guardrails
Predefined rules and limits that prevent an agent from doing wrong or harmful things. Like a speed limiter on a car.
Agent Loop
The cycle an agent repeats: observe, think, act, check. Unlike a chatbot's single-pass response, agents keep looping until the job is done.
Workflow Automation
Using agents to handle repetitive multi-step processes from start to finish. Like an assembly line that moves parts through stages automatically.
Orchestrator
A central agent that coordinates multiple specialist agents. Like a project manager assigning tasks to designers, writers, and developers.
Day 4 - Setup
SOUL.md
The personality file that defines who the agent IS - tone, rules, character.
Workspace Files
SOUL.md + AGENTS.md + USER.md + MEMORY.md = the complete "system prompt" in OpenClaw.
Pipeline
A designed sequence of steps in LangGraph with checkpoints and hard gates.
Human-in-the-Loop
Human approval at key decision points before an agent proceeds.
Sub-Agent
A temporary child agent spawned for one task. Gets the right glossary, does the job, disappears.
Heartbeat
Proactive scheduled check-in - the agent works while you sleep.
Cron Job
A scheduled task at a specific time. Don't survive restarts - recreated from manifests.
Compaction
Summarising long conversations to free up context window space.
pgmemory
PostgreSQL-based long-term memory. Auto-captures, auto-recalls. Not perfect but better than zero.
Atlas Workflow
The practical sequence country managers see: receive task, apply change, run checks, review output, then confirm.
Fallback Chain
Automatic backup routing keeps work moving if a primary model is unavailable.
Glossary Governance
Keeping existing terminology accurate across markets so Atlas uses the right local language every time.
No terms match your search.