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. You start by getting a model to answer you the way you meant, and you finish running an agent unattended that can account for every action it took. 238.7 hours of it is already filmed, so you can be watching the first class ten minutes from now — and the next one runs live on 7 Sep, if you would rather be in the room for it.
Two ways in, both by the month: $250 for the recordings and the written curriculum; $500 for those plus a seat in every class on this path that runs this month. Either one renews monthly and you cancel it yourself, any time — and a class you finish stays yours after you stop.
tms path build-ai-systems --stats
- classes
- 13
- capabilities
- 99
- recorded
- 185 clips / 238.7 h
- exercises
- 625
- student apps
- 13 (1 fully guided)
- tops out at
- order 14 · Paradigmatic
- hours of material
- ~421 h
An autonomous agent with memory.
One that runs without you watching it, remembers what happened last time, and picks up where it left off — behind a review gate, with a bill you can account for.
A laptop and an API key.
That is the whole entry requirement. The first class takes you from there.
The bill nobody models first
What will this thing cost you to run?
The demo is free and the bill is not. Set your traffic and your token mix, put today's prices in the rate fields, and see where a rented GPU stops being the expensive option.
Rates — change any of them
List prices to start with; everything on the right recomputes from whatever you type. Tokens a second means real traffic, not the benchmark chart. A month is 30.44 days.
Your bill, Claude Opus 4.1, 10,000 requests a day
$29,679
$355,875 a year · 9.750¢ per request
Claude Opus 4.1
$29,679 $355,875 a year $15.00 in / $75.00 out per million tokensCheaper model
$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 your token mix and your rates, self-hosting overtakes the cheaper hosted model at about 8,300 requests a day, and beats Opus at about 450 requests a day.
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 bill becomes $20,547 a month, $9,132 off, without changing vendor, model, or anything a user can see. That lever is the one nobody reaches for first, and pulling it is what most of this path is about.
Control AI Spending
Segment text into tokens and cost it. Run an open model on your own machine. The class that makes the number above yours instead of somebody else's.
Context Engineering
Design a recursive summarizer for documents several times the size of the context window — 282 exercises 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.
Keep It Running
Produce an itemized bill for what your system costs, and work out what one user costs you to serve — including the user who costs several times the median.
Already on the calendar
The next 7 sessions on this path
The live plan is a seat in every one of them. Turn up with the traceback you are actually stuck on, ask about it out loud, and take the recording home afterward — it lands in the same login as the 238.7 hours that are already there.
tms schedule --path build-ai-systems
- 7Sep Agentic SDLC Monday · 26.5 h of it already filmed · 46 exercises · $500 on its own next up
- 12Sep Control AI Spending Saturday · 3.1 h of it already filmed · 12 exercises · $100 on its own
- 15Sep Intro to Agents Tuesday · 56.7 h of it already filmed · 32 exercises · $350 on its own
- 16Sep Context Engineering Wednesday · 32.7 h of it already filmed · 282 exercises · $400 on its own
- 17Sep Production Agent Engineering Thursday · 22.3 h of it already filmed · $200 on its own
- 25Sep AI Alignment Friday · 1 exercise · $300 on its own
- 9Oct RAG & Memory Friday · 11.2 h of it already filmed · 47 exercises · $400 on its own
Two ways in
Watch it all, or be in the room
Same curriculum either way. The difference is whether you are asking your questions out loud on the day, starting with Agentic SDLC on 7 September.
$250a month
Watch it all
- 185 clips and 238.7 hours across 13 classes, open the minute you join
- 625 exercises and the written curriculum for all of them
- The 13 tools that come with the classes
- Pause, rewind, and run the exercise with the class on the other screen
$500a month
Be in the room
- Everything on the left, plus a seat in every class on this path that runs this month
- Bring your own traceback and ask about it out loud — 7 sessions are already on the calendar
- Every session is filmed as it runs, so the hour you miss arrives days later
- The archive keeps growing while you are in it
Both renew every month and you cancel either one yourself, any time. A class you finish stays yours — the recordings, the exercises and the written curriculum for it remain in your login after the subscription stops. Classes can also be taken one at a time, at their own prices — this path runs from $100 to $500 a class.
What you walk out able to do
99 things you can prove you can do
Every one of these is a verb you can do afterward, and every one comes with the artifact you hand over to show you can. They stack: the step you are standing on is what lets you reach the next. The last one asks for a week of unattended agent runs and a replay of one of them reconstructed from the trace alone.
- follow a model quickstart model behavior
- locate a tool in an open tool hub system construction
- run a program someone else wrote and read its error system construction
Prove itfollow a model quickstart — A terminal transcript or notebook showing a successful call you ran yourself, including the model's response.
- execute a tool call round trip by hand system construction
- locate a claim in its primary source evidence and verification
- locate the devices a passage uses on you distribution and influence
- run an open model locally infrastructure
Prove itexecute a tool call round trip by hand — A transcript of all four messages — request, tool call, tool result, final answer — from a script with no agent library in it.
- configure an assistants memory and outside connections interaction and interface
- configure an open model endpoint behind a provider interface infrastructure
- configure inference hyperparameters model behavior
- elicit output from a model model behavior
- express a data shape as a schema data shaping
- operate a model as a first pass editor model behavior
- produce a handoff that survives the author leaving organization and delegation
- produce a list of where untrusted input enters a system security
- produce a moderation score for an output alignment
- produce a refusal boundary that fires on cases you did not list model behavior
- produce a reusable system prompt model behavior
- produce a rubric a model can apply evidence and verification
- produce a stopping condition an agent can check on itself model behavior
- produce a usable interface for an agent interaction and interface
- produce an account of the gap between ideals and operating values alignment
- produce an account of what you depend on and who can revoke it infrastructure
- produce an inventory of what you pay for and what it holds infrastructure
- produce an itemised bill for what your running system costs infrastructure
- segment text into tokens and cost it model behavior
- transform a corpus into an embedded index data shaping
- transform a transcript into a structured record data shaping
Prove itconfigure an assistants memory and outside connections — Memory on with one fact it retained across a fresh conversation, one connector enabled with the scopes it was granted written out, and one source you deliberately did not connect with the reason.
- characterise a models failure modes model behavior
- characterise a plans weaknesses with adversarial review evidence and verification
- characterise an applications injection surface security
- characterise how a style works on a reader distribution and influence
- characterise how an ai assisted attack unfolded security
- characterise what a tool costs you beyond its price infrastructure
- characterise what breaks when you cut the connection infrastructure
- characterise what one user costs you to serve infrastructure
- characterise where a teams work stalls organization and delegation
- characterise who gains and who absorbs the cost of a deployment alignment
- classify harmful outputs against a values statement alignment
- classify models by fit for a workload model behavior
- classify where ai can fill a shape and where it must not choose one interaction and interface
- classify which framework applies where you are in the lifecycle security
- constrain model output to a schema model behavior
- generalise a reasoning prompt pattern model behavior
- generalise a specification an agent can build from organization and delegation
- generalise a tool interface for model use system construction
- generalise several sources into a position you can defend evidence and verification
- parameterise a prompt template model behavior
- parameterise an agent persona model behavior
Prove itcharacterise a models failure modes — A table of failure modes, each with the input that reproduces it and the output it produced, reproducible by a second person.
- falsify a claim that a text was written by a person evidence and verification
- falsify a models alignment guardrails alignment
- falsify a prompt with a benchmark evidence and verification
- falsify an applications defences against a named attack class security
- hold several audience models on one surface interaction and interface
- justify a safety case for a system that acts without you security
- justify a vector index for a latency and corpus budget infrastructure
- justify an embedding model and its dimensionality data shaping
- measure a models value alignment against a stated standard alignment
- measure whether a fine tune changed behaviour model behavior
- verify a model output you thought was random model behavior
- verify a program does what you claimed with a test system construction
- verify a rubric against independent graders evidence and verification
- verify an agents actions with a critic evidence and verification
- verify model written code against its specification evidence and verification
- verify synthetic data preserves the property you need data shaping
- verify you would know your system broke before a user tells you infrastructure
Prove itfalsify a claim that a text was written by a person — A judgement on passages of both kinds where you did not know the answer in advance, your score, and the cases you got wrong with the reason.
- construct a multi agent conversation with turn taking system construction
- construct a predictive account of a models behaviour model behavior
- construct a retrieval system system construction
- construct an agent that uses tools system construction
- design a chunking strategy against measured retrieval data shaping
- design a packaged assistant over your documents system construction
- design a recursive summariser for oversized documents data shaping
- design a review gate that catches what the doer cannot see organization and delegation
- design a system you can afford to keep running infrastructure
- design an evaluation regime that decides model changes evidence and verification
- design an interaction protocol for responsible use interaction and interface
- instrument a restore you have actually rehearsed infrastructure
Prove itconstruct a multi agent conversation with turn taking — A transcript where the speaker order differs between two runs of the same task, both terminating, plus the selection rule in source.
- arbitrate between the measure a system optimises and the goal alignment
- arbitrate which internal feature drives a models behaviour alignment
- govern a change to a running system you can undo infrastructure
- govern a swarm by information hierarchy organization and delegation
- re architect an agent as a stateless reducer system construction
- reconcile a failing retrieval by reranking and rewriting data shaping
- reconcile a working set with the context window it must fit model behavior
- reconcile agents from different toolchains into one run system construction
- reconcile an agents run with a tool that failed system construction
- reconcile what you built with what you can maintain alone infrastructure
- select among agent architectures system construction
- select among operating boundaries for inputs a system was not built for alignment
- synthesise a long work that stays consistent across many generations interaction and interface
- synthesise a memory hierarchy for an agent system construction
- synthesise a metalanguage for a problem domain data shaping
- synthesise a retrieval over relationships not just similarity data shaping
- synthesise a self extending agent behind a review gate system construction
- synthesise an accountability regime for an unattended agent system construction
- synthesise an agent that carries notes across its own runs system construction
Prove itarbitrate between the measure a system optimises and the goal — The measure, the outcome it stands in for, and a concrete behaviour that scores well on the first while failing the second.
Tagged by philosophy, this path is 63 think · 61 make · 37 own · 11 influence · 11 secure. A capability can carry more than one tag, so those add up to more than 99.
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 class fills up as you collect the 99 things you can do by the end: purple is what you walked in with, cyan is what that class hands you.
-
01
Using Large Language Models
- synthesise a metalanguage for a problem domain
- reconcile a working set with the context window it must fit
- design an interaction protocol for responsible use
- +11 more
Prove itThe notation's grammar, a work produced in it that exceeds the context window, and a consistency check a reader runs to catch contradictions.
- 6 clips · 10.7 h
- 55 exercises
- Syllabus
-
02
AI Alignment
- originate oversight that holds when the system outpaces the reviewer
- synthesise an accountability regime for an unattended agent
- select among operating boundaries for inputs a system was not built for
- +22 more
Prove itThe scheme, a run where output exceeded what a reviewer could read, and an error the scheme surfaced anyway.
- 1 exercise
- Syllabus
- next session 25 Sep
- $300 on its own
-
03
Context Engineering
- synthesise an agent that carries notes across its own runs
- synthesise a memory hierarchy for an agent
- select among agent architectures
- +36 more
Prove itPaired runs of the same multi-step task with the note store kept and cleared, where the cleared run repeats a step the noted run skips, plus the note text the agent wrote and the later turn that cites it.
- 24 clips · 32.7 h
- 282 exercises
- Guided app
- next session 16 Sep
- $400 on its own
-
04
Claude Model Context Protocol
- generalise several sources into a position you can defend
- generalise a tool interface for model use
- produce a usable interface for an agent
- +3 more
Prove itThe position, the sources behind each claim, and the disagreement between two of them with the reading you chose and why.
- 3 clips · 5.2 h
- Syllabus
-
05
Intro to Agents
- synthesise a self extending agent behind a review gate
- select among agent architectures
- reconcile agents from different toolchains into one run
- +15 more
Prove itA diff the agent authored to its own prompt or toolset, the benchmark run that accepted it, and a logged rejection with the reason recorded.
- 42 clips · 56.7 h
- 32 exercises
- Syllabus
- next session 15 Sep
- $350 on its own
-
06
Control AI Spending
- classify models by fit for a workload
- characterise what one user costs you to serve
- characterise what a tool costs you beyond its price
- +4 more
Prove itA comparison table over at least three models, the same prompt run on each, and the written decision with the property that decided it.
- 2 clips · 3.1 h
- 12 exercises
- Syllabus
- next session 12 Sep
- $100 on its own
-
07
Prompt Engineering
- synthesise an agent that carries notes across its own runs
- synthesise a self extending agent behind a review gate
- synthesise a metalanguage for a problem domain
- +41 more
Prove itPaired runs of the same multi-step task with the note store kept and cleared, where the cleared run repeats a step the noted run skips, plus the note text the agent wrote and the later turn that cites it.
- 58 clips · 63.5 h
- 142 exercises
- Syllabus
-
08
RAG & Memory
- synthesise an agent that carries notes across its own runs
- synthesise a memory hierarchy for an agent
- select among agent architectures
- +17 more
Prove itPaired runs of the same multi-step task with the note store kept and cleared, where the cleared run repeats a step the noted run skips, plus the note text the agent wrote and the later turn that cites it.
- 9 clips · 11.2 h
- 47 exercises
- next session 9 Oct
- $400 on its own
-
09
Advanced Retrieval Augmented Generation
- reconcile a failing retrieval by reranking and rewriting
- design a chunking strategy against measured retrieval
- construct a retrieval system
- +4 more
Prove itA set of questions the first-pass retriever answered wrongly, the reranked or rewritten run, and the before-and-after scores on the same set.
- 5 clips · 6.8 h
- 7 exercises
- Syllabus
-
10
Production Agent Engineering
- originate an operating model for managed agent teams
- synthesise an accountability regime for an unattended agent
- synthesise a self extending agent behind a review gate
- +10 more
Prove itA written operating model in use by at least two people, with a recorded instance of it catching a failure that the previous arrangement missed.
- 19 clips · 22.3 h
- Syllabus
- next session 17 Sep
- $200 on its own
-
11
Agentic SDLC
- synthesise an agent that carries notes across its own runs
- synthesise an accountability regime for an unattended agent
- reconcile what you built with what you can maintain alone
- +11 more
Prove itPaired runs of the same multi-step task with the note store kept and cleared, where the cleared run repeats a step the noted run skips, plus the note text the agent wrote and the later turn that cites it.
- 17 clips · 26.5 h
- 46 exercises
- Companion
- next session 7 Sep
- $500 on its own
-
12
Agentic AI Security: Securing What You Build
- synthesise an accountability regime for an unattended agent
- verify model written code against its specification
- justify a safety case for a system that acts without you
- +4 more
Prove itA week of unattended runs with a per-action trace, and a replay of one run reconstructed from the trace alone.
- 1 exercise
- Reference
-
13
Keep It Running
- reconcile what you built with what you can maintain alone
- govern a change to a running system you can undo
- instrument a restore you have actually rehearsed
- +6 more
Prove itA month you were away with the system still serving users, the runbook somebody else used, and the thing you deliberately shut off because you could not keep it alive.
Straight answers
What is behind the login
238.7 hours of this path, filmed
Prompt Engineering is 58 clips and 63.5 hours; Intro to Agents is 42 clips and 56.7 hours; Context Engineering is 24 clips and 32.7 hours. 185 clips and 625 exercises in all.
Pause it, rewind the part where the trace does not match the code, and run the exercise with the class still on the other screen. Join on a Tuesday and you can be three classes in by the weekend.
And the room they were filmed in
A recording answers the question the teacher expected. The room answers the one you brought about your own repo. The next one is Agentic SDLC on Monday 7 September, and your pass is a seat in it.
Every session is filmed as it runs, so the hour you could not make arrives in the same login a few days later, and the archive you joined keeps getting longer while you are in it.
Both plans, and what carries on
The recordings, the exercises, the tools and the written curriculum come with either one; the dearer one adds the live room. A class you finish stays yours after you stop paying.
Watch it all $250 Be in the room $500The 13 tools that come with the classes
You open these next to your own work — your prompt, your repo, your bill — with the recording paused on the other screen.
- Context Engineering 1 guided workbench. Nine stations that walk you from a bloated prompt to a context budget you can defend, on your own material.
- Agentic SDLC 1 companion, 9 walkthrough decks. A companion for running the loop, plus decks on agent memory, the complexity ladder, context compression, git under agents, and shipping.
- Agentic AI Security: Securing What You Build 1 walkthrough deck, 1 reference. A frameworks reference and a map of where untrusted input gets into the thing you built.
Read this before you pay
Who this path is not for
- 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.
- 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 big. 421 hours of material and 625 exercises; Context Engineering alone has 282 of them. Come for the long haul.
- You want to wander, not to climb. Both plans here are 13 classes deep and one track wide, in a curated order. Somebody who wants to sample across all nine paths should start from the paths index and pick the one they mean.
If you want AI doing your work without building the plumbing yourself, two other paths cover the same ground with no terminal in them:
One sequence, worked out already
Press play tonight, be in the room on 7 September
185 clips and 238.7 hours open the moment you join, in the order the school teaches them, and every class above tells you what you will be able to do — and what you hand over to prove it — before you spend an hour on it. That next session is Agentic SDLC, and the live plan is a seat in it.
Both plans renew every month and you cancel yourself, any time. A class you finish stays yours afterward.