For people who want to ship the system, not read about it
Getting an answer is easy. Getting it to stay up is the job.
Thirteen classes in the order the school teaches them — from your first API call to an unattended agent that carries notes across its own runs behind a review gate. 230.5 hours of it is already recorded, so most of this path is watchable the hour you enrol.
Pay what you can from $100. Every count on this page is measured from the curriculum.
tms path build-ai-systems --stats
- classes
- 13
- capabilities
- 41
- recorded
- 179 clips / 230.5 h
- exercises
- 625
- student apps
- 13 (1 fully guided)
- tops out at
- order 13 · Metasystematic
- est. total load
- ~413 h
has recordings live session only, no video yet — 4 of 13
The bill nobody models first
What will this thing cost you to run?
Step two of this path is Control AI Spending, and it is step two for a reason: the demo is free and the bill is not. Set your traffic and your token mix, and see where a rented GPU stops being the expensive option.
These rates are illustrative. They are round numbers in the published range for hosted inference and rented accelerators, used here so the arithmetic is inspectable. They are not The Multiverse School's prices, they are not any vendor's committed pricing, and they move. Self-hosting is modelled as one 80 GB GPU at $1.80/hour doing 6,000 prefill and 700 decode tokens a second at 60% duty — plausible, not promised. A month is 30.44 days. Put your own provider's numbers in before you budget anything.
Your bill, mid-tier API, 10,000 requests a day
$1,583
$18,980 a year · 0.520¢ per request
Frontier API
$29,679 $355,875 a year $15.00 in / $75.00 out per million tokensMid-tier API
$1,583 $18,980 a year $0.80 in / $4.00 out per million tokensSelf-hosted
$1,315 $15,768 a year 1 rented GPU at $1.80/hour, running all monthA rented GPU costs the same whether you serve ten requests or ten thousand. At this token mix it overtakes the mid-tier API at about 8,300 requests a day, and it beats the frontier API at about 450.
You are at 10,000 a day, so self-hosting is the cheaper option here — and you have just taken on a GPU to keep alive.
Halve the prompt and the mid-tier bill becomes $1,096 a month — $487 saved without changing vendor, model, or anything a user can see. That lever is the one nobody reaches for first.
Control AI Spending
Segment text into tokens and cost it; run an open model locally. 2 clips / 3.1 h recorded, 12 exercises. The class that makes the number above yours instead of mine.
Context Engineering
Design a recursive summariser for oversized documents. 27 clips / 35.7 h and 329 exercises — the largest exercise set on the path, aimed squarely at the input-token half of that bill.
Production Agent Engineering
Configure an open model endpoint behind a provider interface — so switching between hosted and self-hosted is a config change rather than a rewrite. 19 clips / 22.3 h.
Keep It Running
Produce an itemised bill for what your running system costs, and characterise what one user costs you to serve. No recording yet — this one is the live session and its 5 capabilities.
The sequence
Thirteen classes, from prompt to deployed system
This order is curated, not alphabetical and not chronological — each class stands on the one above it. The bar under each row is the recorded archive filling up as you go: 230.5 hours by the end, with the cyan segment showing what that class adds.
-
01
Using Large Language Models
Leaves you able to generalise a reasoning prompt pattern, plus 4 more capabilities.
- 6 clips · 10.7 h
- 55 exercises
- No app yet
- 5 capabilities
-
02
Control AI Spending
Leaves you able to segment text into tokens and cost it, plus 1 more capability.
- 2 clips · 3.1 h
- 12 exercises
- Syllabus
- 2 capabilities
-
03
AI Alignment
Leaves you able to measure whether a fine tune changed behaviour, plus 7 more capabilities.
- no recording yet
- 1 exercise
- No app yet
- 8 capabilities
-
04
Context Engineering
Leaves you able to design a recursive summariser for oversized documents, plus 2 more capabilities.
- 27 clips · 35.7 h
- 329 exercises
- Guided app
- 3 capabilities
- live 16 Sep
-
05
Claude Model Context Protocol
Leaves you able to generalise a tool interface for model use, plus 2 more capabilities.
- 3 clips · 5.2 h
- Syllabus
- 3 capabilities
-
06
Intro to Agents
Leaves you able to construct an agent that uses tools, plus 4 more capabilities.
- 42 clips · 56.7 h
- 32 exercises
- Syllabus
- 5 capabilities
- live 15 Sep
-
07
Prompt Engineering
Leaves you able to synthesise a metalanguage for a problem domain, plus 5 more capabilities.
- 58 clips · 63.5 h
- 142 exercises
- Syllabus
- 6 capabilities
-
08
RAG & Memory
Leaves you able to synthesise an agent that carries notes across its own runs, plus 2 more capabilities.
- no recording yet
- No app yet
- 3 capabilities
-
09
Advanced Retrieval Augmented Generation
Leaves you able to construct a retrieval system, plus 1 more capability.
- 5 clips · 6.8 h
- 7 exercises
- No app yet
- 2 capabilities
-
10
Production Agent Engineering
Leaves you able to reconcile agents from different toolchains into one run, plus 5 more capabilities.
- 19 clips · 22.3 h
- No app yet
- 6 capabilities
- live 17 Sep
-
11
Agentic SDLC
Leaves you able to synthesise an agent that carries notes across its own runs, plus 6 more capabilities.
- 17 clips · 26.5 h
- 46 exercises
- Companion
- 7 capabilities
- live 7 Sep
-
12
Agentic AI Security: Securing What You Build
Leaves you able to verify model written code against its specification, plus 2 more capabilities.
- no recording yet
- 1 exercise
- Reference
- 3 capabilities
-
13
Keep It Running
Leaves you able to verify you would know your system broke before a user tells you, plus 4 more capabilities.
- no recording yet
- No app yet
- 5 capabilities
Straight answers
What is actually behind the login
The recordings — 9 of 13 classes
179 clips, 230.5 hours, 625 exercises. This is the deepest archive of any path in the catalog, and it is the reason this is the one page where "watch it on your own schedule" is the whole truth rather than a consolation.
The three heaviest are Prompt Engineering (58 clips / 63.5 h), Intro to Agents (42 / 56.7 h) and Context Engineering (27 / 35.7 h). Those three alone are 155.9 hours.
4 classes have no recording at all
AI Alignment, RAG & Memory, Agentic AI Security: Securing What You Build, Keep It Running. For those you get the live session, whatever exercises exist, and the reference material — not a video library. 0 of the 4 is on the calendar right now; the rest run when they are next scheduled.
AI Alignment is the sharp one: 8 capabilities, more than any other class on this path, and none of it is on tape. If you are buying this path for the archive, buy it for the other nine classes and treat that hour as a live hour.
6 classes have no app
The app rung measures how far you get on your own with no one in the room. 6 classes here are at zero — Using Large Language Models, AI Alignment, RAG & Memory, Advanced Retrieval Augmented Generation, Production Agent Engineering, Keep It Running — which means the class is the session and the recordings, with nothing that walks you through it afterwards.
The thinnest by that measure is Advanced Retrieval Augmented Generation: no app, and 5 clips against a topic that deserves more. We would rather you know that now.
The live hour
The hour behind this path is Make & Do — Mon-Thu, 9:00 Pacific, led by neek — the standing session of the Shipping Software programme. It is where the classes above get applied to whatever you are actually building that week.
4 of the 13 classes have a live session on the calendar right now. The other 9 are available as whatever they have recorded, and run again when they are next scheduled.
What the 413 hours means
That figure is recorded video, plus two hours per class for the live session, plus fifteen minutes per exercise. The first two are measured. The fifteen minutes is a guess, and it is the only number on this page that is — 625 exercises is a lot of guesswork to stack up, so treat 413 as an order of magnitude, not a promise.
That is the whole inventory
179 clips, 625 exercises, 13 apps and four classes that are a live hour and nothing else. If that is worth $250 to you, it is the Video tier.
Enrol $250The 13 apps, by name and link
Three of the thirteen classes ship something you can open on your own. The other ten do not — which is the honest reading of the list below, not a rounding error.
Context Engineering 1
- /x/context-workbench wizard
Agentic SDLC 10
- /x/agentic-sdlc-companion companion
- /x/agentic-sdlc-agent-memory deck
- /x/agentic-sdlc-complexity-ladder deck
- /x/agentic-sdlc-compression deck
- /x/agentic-sdlc-cto-mindset deck
- /x/agentic-sdlc-devops-git deck
- /x/agentic-sdlc-devops-ship deck
- /x/agentic-sdlc-git-situations deck
- /x/agentic-sdlc-outdoor-pet deck
- /x/agentic-sdlc-probabilistic deck
Agentic AI Security: Securing What You Build 2
One is fully guided — the Context Engineering workbench, the only rung-5 app on the path. The rest are companions, decks and references: useful next to the class, not instead of it.
The 41 capabilities
How far up this actually goes
Every capability on this path, by complexity order. It starts at following a quickstart and ends at order 13 — Metasystematic: reconciling agents from different toolchains into one run. Nothing else in the catalog goes higher.
- system construction 12
- model behavior 11
- infrastructure 7
- evidence and verification 5
- data shaping 3
- security 2
- organization and delegation 1
- locate a tool in an open tool hub system construction
- follow a model quickstart model behavior
- run an open model locally infrastructure
- execute a tool call round trip by hand system construction
- produce a list of where untrusted input enters a system security
- configure inference hyperparameters model behavior
- produce a rubric a model can apply evidence and verification
- produce an account of what you depend on and who can revoke it infrastructure
- produce an itemised bill for what your running system costs infrastructure
- transform a corpus into an embedded index data shaping
- operate a model as a first pass editor model behavior
- produce a reusable system prompt model behavior
- elicit output from a model model behavior
- configure an open model endpoint behind a provider interface infrastructure
- segment text into tokens and cost it model behavior
- characterise an applications injection surface security
- generalise a tool interface for model use system construction
- characterise what one user costs you to serve infrastructure
- generalise a reasoning prompt pattern model behavior
- characterise what breaks when you cut the connection infrastructure
- constrain model output to a schema model behavior
- parameterise a prompt template model behavior
- characterise a models failure modes model behavior
- verify model written code against its specification evidence and verification
- verify an agents actions with a critic evidence and verification
- measure whether a fine tune changed behaviour model behavior
- verify you would know your system broke before a user tells you infrastructure
- verify a rubric against independent graders evidence and verification
- verify a program does what you claimed with a test system construction
- falsify a prompt with a benchmark evidence and verification
- design a recursive summariser for oversized documents data shaping
- construct an agent that uses tools system construction
- construct a retrieval system system construction
- design a review gate that catches what the doer cannot see organization and delegation
- construct a multi agent conversation with turn taking system construction
- synthesise an agent that carries notes across its own runs system construction
- select among agent architectures system construction
- synthesise an accountability regime for an unattended agent system construction
- reconcile agents from different toolchains into one run system construction
- synthesise a self extending agent behind a review gate system construction
- synthesise a metalanguage for a problem domain data shaping
Tagged by philosophy, this path is 32 make · 22 think · 14 own · 3 secure. A capability can carry more than one tag, so those add up to more than 41.
Read this before you pay
Who this path is not for
- It assumes you will write code and live in a terminal. Not "a bit of Python eventually" — from Intro to Agents onward you are running processes, reading stack traces and configuring endpoints yourself.
- It assumes you want to own the running system, including the bill, the outage and the injection surface. If you want the output and not the operations, this is the wrong door.
- It is long. 413 estimated hours, 625 exercises. There is no version of this that takes a weekend.
If you want AI doing your work without you building the plumbing, two other paths cover the same ground with no terminal in them:
One sequence, worked out already
Start at the first class and stop thinking about the order
179 clips and 230.5 hours are waiting the moment you enrol. The four classes without recordings run live, and the sequence above tells you exactly what you are getting from each one before you spend an hour on it.
Pay what you can from $100. Or take any single class for $60.