Automating the Boring Parts Without Losing Control
Bring a COPY of a messy folder — say it twice. The demo is on a copy.
60-second recap
The deal
AI for friction, humans for meaning. Raw output is never finished work. Verify before trusting. Disclose where it matters. Skepticism welcome when specific.
Standing recap.
The shift
Last week it wrote a file. This week it acts.
An agent breaks a goal into steps, uses tools, reads and writes files, checks itself, continues.
An extremely fast intern with no judgment, no stake, total confidence.
"Used well: your evenings back. Used carelessly: it renames your only copy at 2am and feels great about it."
The week in one line
Automate friction, never judgment — and put a human at every door that doesn't open backwards.
Board: PLAN FIRST · CHECKPOINT · RAILS · CHANGELOG.
Deliverable
You will leave with
An agent workflow spec — steps, tools, risks, checkpoints An Automation Boundary Map — automate / assist / human-only
Both are design documents. No agent software required to build them.
Today's map
Today's map
Agents, tool use, checkpoints, rails — 40 min
The Automation Boundary Map — 15 min
Demo: an agent organizes a messy folder (supervised) — 35 min
Lab: your boundary map + workflow spec — 70 min
Share: find the missing checkpoint — 35 min
Breaks protected.
Part 1
Agent vs chatbot vs workflow
~0:10.
Three things, not one
Sort them
Chatbot — answers; you drive each turn; low risk
Workflow — fixed steps you defined; you wrote the path
Agent — picks its own steps; higher risk
Most students conflate all three. This sorts them.
The reframe
You don't need a robot.
Most "agentic" creative work is a workflow — predictable, repeatable, safe — with one or two agentic steps.
You need an assembly line with a human inspector.
Anthropic's own engineers: prefer the simplest pattern that works.
Part 1, continued
Tool use
Where capability and danger both live
~0:22.
Every tool is a risk surface
Smallest access that does the job
An agent's power = tool use: read/write files, run code, search, call services.
"A folder-organizer needs to move files. It does not need the internet or the power to delete."
Why scope tightly
Broad access is how
"organize my folder"
becomes "email my drafts to a client."
Concrete, memorable, true. Scope is a safety decision.
Part 1, the thesis
The checkpoint
Where you stay the director
~0:30.
The pattern
Plan first, act second
Make the agent propose its plan — and STOP for approval before touching anything.
Two words: plan first. The whole safety pattern compresses to this.
No agent software needed. The spec is the deliverable.
Parts + times on board.
Lab · step 1
15 min
Boundary map
10–15 real tasks → Automate / Assist / Human-only.
Defend your "human-only" column to a neighbour.
Empty human-only column → "Really? The concept? The client trust?"
Lab · step 2
30 min
Pick + spec
One "automate" task → template completely: steps · smallest access per step · a checkpoint before every irreversible action · the four rails · a verification method.
✏️ each checkpoint with what would make you reject.
The checkpoint-before-irreversible is the assessed element.
Lab · step 3
15 min
Critique swap
Trade specs. Find in your partner's: the missing checkpoint and the over-broad access.
Every first draft has at least one.
This peer-audit is the real teaching of the day.
Always available
Alternative
Zero-AI / paper track (fully equal)
The boundary map and the spec are design documents.
A workflow you can't write on paper, you can't trust to a machine.
This week's lab is paper-native. Say so.
Share-out · 35 min
Find the missing checkpoint
Read the chore, the irreversible step, the checkpoint that guards it.
Room asks: "What happens if that checkpoint isn't there?"
One "human-only" column read aloud — applaud the refusals.
Every week
Integrity check
This week's four
A checkpoint before EVERY irreversible step?
Any step with broader access than the task needs?
Whose material does it touch — all permission-cleared?
Client deliverables: "organized by a supervised agent; all client-facing decisions human"
Four beats.
Reflection
Write 4 lines
The task you moved from "automate" to "assist" once you thought about reversibility
The checkpoint you originally forgot
One thing in "human-only" you'll never let an agent near, and why
Keep the spec — a supervised workflow is a strong capstone shape
Two quiet minutes. Plant the Week 14 seed.
Next Wednesday
Jobs, Portfolios, and Career Leverage
Turning everything you've built into the materials that get you hired — in your voice, not generic sludge.
Bring a real project you'd put in a portfolio.
End on time; corridor questions after.
Tonight, optional
Go deeper
Anthropic — Building Effective AI Agents (this week's stance, from the source)
Anthropic — Writing Tools for AI Agents (scope and least-privilege)
The EU AI Act explorer — automated decisions are tightening
Design a second workflow for a collaborative chore