You're about to teach people how to use The AI Surfer OS. This is the shared ground under all four of you: what the thing is, the words we use, what makes an agent good, and the four skills you'll hand over. Written to be taught from, not just read.
It's a t-shirt that goes over Hermes.
Underneath the OS there's an agent runtime — usually Hermes. That's the engine. It talks to the AI model, runs tools, keeps sessions alive. Powerful, and not something a business owner ever wants to touch.
The OS is the layer on top that makes it usable: the dashboard, the crew with names and faces, the board, the schedules, the Beach. Your students never open a terminal. They tap things and talk to their team.
Say this early and say it plainly: we didn't build the AI. We built everything around it that makes it feel like a team you employ. That one sentence prevents a lot of confused questions later.
Not a chatbot with a name. An agent is a worker with an identity, a memory, a set of tools, and permission to act.
Who they are and how they work — their role, their voice, their standing instructions. Their character sheet.
SOUL.mdWhat the boss calls them, what they're responsible for, which department they sit in.
profile.yamlFacts they keep. Their own private notes, plus a whiteboard the whole crew shares.
REMEMBER: · TEAM:What they can physically do — files, terminal, browser, and every app you connect.
skills · ComposioVerbs that change the business: file a task, hand off, schedule, deliver, hire.
TASK: · HANDOFF: · CRON:What they know this turn — the business, the board, the recent conversation.
Business InfoGet these right and everything else is teachable. Get them wrong and every later explanation wobbles.
The boss says something, or a schedule fires.
The agent is handed its instructions, memory, the board, the conversation.
Thinks, uses a tool, reads the result, uses another. Often many times over.
Files a card, saves a file, hands work to someone.
One answer comes back. The turn is done.
Step 3 is why a turn can take ninety seconds and involve a dozen tool calls while the student sees only "thinking". Teach that the wait is work, not lag.
Models don't read letters, they read tokens — chunks of about four characters. "Hello there" is roughly three. Every turn, the agent re-reads its whole briefing, and you pay for all of it again.
Characters, measured on a live box. That's roughly 34,000 tokens per turn before the agent says a word. Useful for you to know. Mention it to students only if they ask why it costs what it costs.
This is the question you'll be asked most, so it's worth being able to answer it cleanly. An agent has no memory of its own between turns. Every turn it reads a fresh briefing cold. So "does it remember?" really means "did that make it into the briefing, or can it go and find it?"
Saved facts that are relevant to what you just said, the last 30 messages of this thread, and the business info. Always there, no asking needed.
the briefingOlder facts, the crew whiteboard, and past conversations — including ones far older than this thread shows. It searches when a question needs it.
RECALLA passing detail nobody kept. It may still be findable in an old conversation, but it was never turned into something the agent knows.
say "remember this"Its instructions, its memory, the board, recent messages. Rebuilt fresh every single turn.
Say: "It's the desk they sit down to. We set the desk every time."
Reading a file, searching the web, sending an email through a connected app.
Say: "Thinking is free. Doing is a tool call."
The receiving agent actually wakes up, does the work, and reports back.
Say: "Like walking a job over to a colleague's desk."
You can have several with the same agent. Each keeps its own history.
Say: "Separate chats, like separate email threads."
The Lineup. Cards move through stages, and agents put cards on it themselves.
Say: "If it's not on the board, it isn't really happening."
GPT, Claude and friends. Swappable — the agent's identity doesn't change with it.
Say: "Same employee, different amount of caffeine."
Most training skips this, and it's the part that decides whether someone still trusts their crew in a month. A good agent isn't the cleverest one. It's the honest one.
No "I've sent that" when it drafted. No "I've fixed it" when it tried. Past tense is a promise. Nothing loses a business owner faster than a confident agent that was wrong.
A good agent that hits a wall tells you out loud. A bad one quietly does something adjacent — which reads exactly like being ignored.
Cards move when work moves. Finished work gets attached. Stale cards are worse than no cards, because they teach the owner not to believe the board.
Money out, messages to real people, publishing, deleting. A good agent prepares freely and pauses at the door.
Durable facts about the business and the boss's preferences — not every passing detail. Memory is for what should still be true next month.
A specialist that hands off is worth more than a generalist that guesses. Departments exist so the right desk gets the job.
The OS runs on a server the owner rents — not on their laptop. Lead with that, because it explains almost every question that follows.
A rented computer that never sleeps. This is why the crew keeps working when the laptop is shut.
Sets up the engine and the OS together, then hands back a web address.
Ties the box to their account and lets it receive updates.
The crew needs a brain to rent. Nothing works until this is done.
Ten minutes here makes every agent dramatically better. Don't let them skip it.
One agent, one real job, finished end to end. That first win is the onboarding.
An agent with no connections can think and write. An agent with connections can run the business. There are two doors in.
One service covering a thousand-plus apps: Gmail, Slack, calendars, CRMs, sheets. The owner logs into the app the normal way and Composio holds the keys. Nobody copies secrets around.
After connecting, tools have to be reloaded before the crew can see them. This step gets forgotten constantly.
For services Composio doesn't cover, or where the owner wants to use their own key. Paste it into Tool Shed, or straight into chat — the OS spots it, files it away, and wires it in for the crew.
Keys stay on their own box. Agents use them by name and never print them.
A scheduled job is an agent doing a turn while nobody's watching. Same worker, same tools, no audience.
The owner asks in plain language — "check my site every morning and tell me if it's down" — and the agent sets it up. Or they pick one from the suggested jobs. Either way it appears on the Cron Jobs page, owned by an agent.
Into the chat with the agent that owns it, and into the Morning Brew. A job that runs and reports nowhere is the most common disappointment — so always check the first run landed somewhere visible.
The owner describes what they want. An agent builds it. They look at it, say what's wrong, it changes. No code, no deploy pipeline, no jargon.
"A booking page for my studio, dark, with my logo."
The agent makes it in one decisive pass and saves it as a real file.
It renders the page on desktop and phone and judges its own work.
"Bigger headline, form higher up." It edits the same page — no second copy.
Goes live on a real address when the owner's happy.
The bread and butter. Edit live on a canvas with the agent beside you.
Small tools that kill one annoying job — a calculator, a quote builder, an internal form.
Tools that talk to connected services, using keys already on the box.
Agents write these themselves — students never type them. Keep this open so you can explain what just appeared on screen.
| Written by an agent | What the owner sees happen |
|---|---|
TASK: | A card appears on the board, assigned to someone |
HANDOFF: | Another agent wakes up and starts working |
ASSIGN: MOVE: | A card changes owner, or moves stage |
DELIVER: | Finished work gets attached to its card |
ARCHIVE: | A stale card is tucked away |
CRON: | A scheduled job is created |
ASK: | Tappable buttons appear for a decision |
HIRE: | A new teammate is proposed |
REMEMBER: TEAM: | A fact is saved privately, or to the crew whiteboard |
RECALL: | The agent searches its own notes and past conversations, then answers with what it found |
REVIEW: | A built page is rendered and judged on desktop and phone |
PROJECT: FOCUS: | An initiative is opened, or the day's focus is set |
Correct these on day one and you'll save yourself weeks.
| They think | Actually |
|---|---|
| "The agents are always awake watching things." | They act when asked, or when scheduled. Nothing happens in between. |
| "If I rename it, it becomes a different agent." | The name is a label. Behaviour comes from its character sheet. |
| "More instructions make it obey better." | Usually the opposite. Clearer role, fewer jobs, shorter instructions. |
| "It remembers everything we've ever said." | It carries saved facts plus the recent conversation, and it can go search the rest when it needs to. If something matters, say "remember this" — that's what makes it stick. |
| "It said it sent the email, so it sent it." | Some actions wait for a tap. Look for the receipt, not the sentence. |
| "It's broken" — when an agent explains a failure | An agent's explanation is a lead, not a finding. Check the thing itself. |
| "One agent can do everything." | It can. It shouldn't. Specialists that hand off beat one generalist. |
Enter the four-digit code to open it.
You'll only need this once on this device. If you don't have the code, ask Zach.