# Session 01 — Wake an Agent · Student Handout

Your name: ________________   Partner: ________________

Tonight's finish line: **an agent that ran one real job for you and is
scheduled to run again tonight while you sleep.** Keep this handout — the
JOB.md you write here is the seed of everything in weeks 2–4.

---

## Exercise 1 · First contact  ·  15 min · solo

**Goal:** see a tool call and a permission prompt with your own eyes.

1. Open your terminal, go to your course folder, start the agent:
   ```
   cd ~/agent-class && claude
   ```
2. Ask it something a chatbot can't answer about your machine:
   ```
   what files are in my Documents folder?
   ```
3. A **permission prompt** appears asking to run a command. **Read it out
   loud.** Then approve it.
4. Notice what happened: you gave a goal in English, it chose a tool, it
   asked permission, it reported back.

**Done when:** you've seen one tool call and one permission prompt.

Write the exact wording of the permission prompt you saw:

`____________________________________________________________`

Why do you think it asked instead of just doing it? ______________________

`____________________________________________________________`

---

## Exercise 2 · Give it hands  ·  30 min · solo (helpers roaming)

**Goal:** your agent sends a message to your phone.

1. Connect the **calendar** connector.
   `[INSTRUCTOR: paste the exact connect steps + screenshots here for
   your class's setup.]`
2. Connect **Hermes** so the agent can message your phone.
   `[INSTRUCTOR: paste the exact connect steps here.]`
3. Ask:
   ```
   read my next three calendar events and send me a one-line summary
   of each, on my phone
   ```
4. Wait for the buzz.

**Done when:** your phone shows a message your agent wrote.

That feeling when your phone buzzed — one word for it: ___________________

---

## Exercise 3 · The standing instruction  ·  25 min · pairs

**Goal:** write a real job brief and run it once, supervised.

Pick a real, recurring, **read-only** chore. Ideas: "summarize my
unread newsletters," "list everything on my calendar this weekend,"
"tell me which of my subscriptions renew this month."

**Fill this in** (copy from `templates/JOB-template.md`):

```
# JOB.md

## GOAL
(one sentence — what should be true when this is done?)

## STEPS
(the rough shape — 3 to 5 bullets, not a script)

## BOUNDARIES
May read: ______________________________
May write: ______________________________
Must NEVER touch: ______________________  (tonight: sending & deleting)

## DONE LOOKS LIKE
(the exact artifact — "a message on my phone with 3 bullets by 6:30 AM")

## IF STUCK
Stop and report. Never guess on anything irreversible.
```

**Before you run it,** your partner asks you: *"what will your agent do
if ___?"* until your brief has an answer. Write two questions they asked
that your first draft couldn't answer:

1. `____________________________________________________________`
2. `____________________________________________________________`

Now run it:
```
read JOB.md and do the job now
```

Where did it surprise you? That spot = a loose line in your brief. Tighten
one line and run again.

**Done when:** one clean supervised run against a tightened brief.

---

## Exercise 4 · Set the alarm  ·  15 min · solo

**Goal:** schedule tonight's run, then close the laptop.

1. Schedule it:
   ```
   run JOB.md every night at 2:00 AM and message me the result when done
   ```
2. Confirm the schedule exists — ask the agent to list your scheduled
   tasks and read it back to you.
3. **The leash check.** Say your never-list to your partner out loud:
   "tonight my agent will never ______, never ______, never ______."
4. Close the laptop.
   `[INSTRUCTOR: note here if the machine must stay awake/plugged in for
   the schedule to fire on your class runtime.]`

**Done when:** schedule confirmed, never-list spoken, laptop shut.

---

## Homework · sleep on it

- Your agent runs at 2 AM. **Do not watch it.** (Watching is a chatbot
  habit — you're breaking it.)
- Tomorrow, read the log and result over coffee. Fill this in:

**One thing it did well:** ___________________________________________

**One thing it fumbled:** ___________________________________________

**Bring both to Session 02.** If it failed completely, bring the failure
log — those are the best ones.

---

### The six parts of every agent (reference)

| Part | What it is |
|---|---|
| 1 · Brain | The model. Rented by the token. The least interesting part. |
| 2 · Hands | Tools: email, calendar, files, browser, messaging. |
| 3 · Memory | Files it reads at start, writes as it learns. |
| 4 · Job | The written standing instruction (your JOB.md). |
| 5 · Clock | The schedule or trigger that wakes it. |
| 6 · Leash | Permissions, approvals, logs. The professional's obsession. |

**Rule zero:** the agent drafts, holds, and asks. It does not send, buy,
or delete. (We loosen this on purpose in Session 04 — never by accident.)