Not just one example. A whole gallery of offices, already running.
Every office in this course is real — built, run, and checked
against real output. Here's the full range, and a few real conversations where one got built live.
33
offices, built & running
9
real build conversations
20
sources & sinks, no code
Find yours
Which team is like yours?
Four shapes cover almost everything. Skim these, find the one
closest to what you have in mind, and describe yours from there.
Watch one thing, speak up
One stream of information comes in; a worker or two watch it and report when something matters.
e.g. room monitor, weather & stock alerts
Take turns, in rounds
Workers go back and forth in rounds, repeating until someone decides it's settled.
e.g. a debate panel
Share one ledger, one at a time
Several workers read and update one shared record; things are handled one item at a time so it stays consistent.
e.g. investment club, customer returns
Propose, then get approval
A worker proposes something and waits for an answer, checked against a shared limit or record, before acting.
e.g. trading desk, loan desk
Already running
Runs on any laptop. No keys, no signups.
Zero or one AI call per cycle — install and see output in seconds.
periodic_brief
Morning brief: news + weather + stocks, one page.
weather_monitor
Hourly plain-English weather briefing.
stocks_monitor
One-line read of a ticker's movement.
loudness_monitor
Mic audio → threshold-crossing alert.
backyard_birds
Audio clips → species ID via BirdNET.
wildlife_watcher
Camera-trap photos → species labels.
recovery_demo
Monte Carlo π — crash, resume, no lost work.
Already running
The richer ones. Real judgment, real reach.
Many AI calls per cycle, or talk to Gmail/Calendar/Slack/your CRM.
periodic_brief_pro
Brief + tagged news + calendar + Gmail.
situation_room
Feeds → four thinkers → synced brief.
situation_room_pro
Same, Claude writes, open-weight extracts.
arxiv_radar
Daily papers → rated → high-relevance digest.
kalshi_market_watch
Prediction markets → briefing on price moves.
competitor_watch
Tech news → entities/sentiment → daily digest.
inbox_triage
Gmail → urgency + summary → Slack digest.
ticket_router
Webhook tickets → classified → oncall Slack.
lead_qualifier
Webhook leads → summarized → CRM.
new_grad_jobs
HN hiring thread → screened for new-grad fit.
job_hunter
RSS jobs → matched to your resume → cover letter.
wardrobe_assistant
Calendar + weather + closet → outfit pick.
For the curious
One idea at a time. Patterns to read.
Small, deliberately minimal offices — each one teaches a single moving part.
my_first_office
Single-agent Hacker News briefer.
org_news_editorial
Two-agent chain — analyst feeds editor.
org_two_office_news
An office made of two other offices.
org_intelligence_briefing
Multi-source feed, significance filter.
org_news_filter
A pipeline that drops non-matches.
org_situation_room
The simpler, two-agent situation room.
webhook_listener
Listens on localhost for HTTP POSTs.
web_monitor
Page-watching via the MCP web source.
gmail_monitor
Pulls unread mail from Gmail.
Now watch one happen
These aren't demos. They're real conversations.
Every case below is an unedited transcript: a fresh assistant, given nothing but the same
instructions and gallery you'll get, handed one plain-English description and left to build. Some are the
first draft; some are the correction; two were carried all the way through to a real, running office.
Nine real cases
Nine offices, nine conversations.
build
Loan desk
Score each application, then check a running risk limit before offering — kept as one worker's own memory.
correction
Loan desk, two desks
Now two desks share one limit — the private total becomes a shared keeper both check before committing.
build
Fractions tutor
One tutor worker holds a kid's mastery and picks the next question — easier, harder, or a new skill.
build
Adaptive tutor
A coach, a quizmaster, a question shelf, a progress keeper — explained back in three plain-English parts.
regression
Adaptive tutor, re-run
Same office, a tougher edit to the instructions — comes out even cleaner the second time.
build
Customer returns
One team handles every customer at once, tagged not duplicated — plus a manager who checks in on any one.
build
News subscriptions
Friends each ask for a different slice of the same story — computed once, shared, not re-billed per person.
built & run
Shipment release
A real first-draft miss, caught and fixed — then actually generated and run, not just discussed.
built & run
Returns desk
Correct on the first try — an urgent ticket freezes the line until the manager answers, then resumes.
Spotlight
what changed, and why it matters
"Now you have two offer-writers, consumer and business, and they share one
limit for the whole desk. If each kept its own private total, neither would ever see the other's committed
loans... So the running total can't live inside the offer-writers anymore. The fix is the same shape as a
trading desk's ledger: move the shared total out into one worker that both offer-writers check with."
— the assistant's own explanation, unedited
One sentence in Pat's
correction ("they both draw against the same limit") and the assistant correctly promoted private memory
into a shared keeper — the exact inverse of the office it had just built.
Spotlight
explaining it back to Pat
"A coach worker is the brain. It remembers how well your kid knows each skill, decides what
to practice next — easier when they're struggling, harder as they improve, and slipping in a bit of review
of older or shaky skills so they don't fade — and only moves on to a new skill once the current one and the
skills it builds on are solid. A quizmaster worker runs the actual practice: shows one
question, waits, checks the answer, offers a gentle hint and another try if it's wrong."
— the assistant's "meet the team," unedited
A fresh assistant, given
only the instructions and the gallery, built this from one paragraph — and correctly avoided a common trap
(no gate on the progress record, because only one worker ever touches it).
Spotlight
a real miss, caught and fixed — then actually run
"I want an office that releases a shipment only after both its warehouse scan and its manifest paperwork have
come in. Match them up by shipment ID, then send the release notice to the loading dock."
— Pat's request, verbatim
The first draft matched scans to manifests correctly — but this
case was carried all the way through: approved, generated, and actually run, not just discussed. Unlike the
others here, you can check its real output for yourself, the same way you'd check any office in this
gallery.
See it yourself.
Every app, every example, and every full cold-test transcript quoted here lives in the
repos, unedited.