What you'll learn
Example
Diagram
Module essentials
Module quiz
Module completed
Operator note
This is where the difference is taught.
Maturity diagnostic, 4D framework, and 3 exportable tools. In one hour you have a defensible adoption plan.
Outcome · when you finish
Pilot path · 60 min
If you're just starting and want to land fast, this is the order. Each step takes you to the recommended content:
When you close the hour: you have a diagnostic, a framework, a governance baseline, a use case modeled in an exportable Tree, and its economic viability number. Enough for an adoption conversation with your team on Monday.
Designed to move from individual exploration to real team use cases.
Click any module to start there. Completed ones are marked.
Common Foundations
Tools — Apply what you learned
Executable diagnostics, calculators, and structurers. Not theory modules — tools where you produce or calculate something concrete.
Track orienter
.
.
/100
/100
| /100 |
Before turning an AI agent loose on your project, it's worth writing down what counts as success and what doesn't.
This tool walks you through eight steps: you define three levels (strategic, tactical, operational),
the real tensions between them, and how they get resolved. At the end you get a CLAUDE.md
ready to paste into the agent.
Give the project a concrete name, not a generic label. "Internal legal assistant — Q3 pilot" says a lot more than "legal chatbot." And tell us why you're kicking it off now.
At least 3 words. Make it clear what it does, not what category it falls into.
At least 20 characters. Write down what changed or what's no longer sustainable.
Who, at the very top, decides whether this was a success? Their intention is the one that, if not met, makes everything else meaningless. The invalidation test is an example where execution was perfect but the intention was violated.
The person who embodies this level and answers if the intention isn't met.
The goal whose fulfillment justifies the project. Just one, as concrete as possible.
Describe a situation where execution went perfectly but the intention got violated. If you can't describe one, it's not a real intention — just a target.
Who coordinates day-to-day, sprint by sprint, week by week? This person has their own idea of success, different from the strategic level — and just as legitimate.
The person who embodies this level.
The purpose that defines success at this level.
A scenario where deliverables were met but the tactical intention got wrecked.
Who does the actual grunt work, hands in the mud? Their constraints are real and legitimate, even though the levels above often don't see them.
The person who embodies this level.
The purpose that defines success at this level.
A scenario where nominally everything looks fine but operations become unworkable.
The three levels, when they cross paths, almost always ask for things that don't fit together. Write down where those real collisions show up in your project. If you can't find any tension, you didn't look hard enough: in practice, saying "no conflict between levels" means you haven't been honest about the project yet.
For each tension you wrote down, decide which intention wins and spell out — in full — what gets sacrificed. If nobody loses anything, you didn't actually decide: you left the call to chance or to whoever pushes hardest later.
Here you set the agent's limits: what it can never do alone, when it has to stop and ask a human, and how someone from outside can verify the result is good (the three-sentence test).
Absolute red lines. Actions that don't happen even if the agent "thinks" it has a good reason.
Situations where the agent can proceed, but only after a human approves.
Write three verifiable statements that someone outside the project can check to decide whether the result is good — without having to ask you anything.
CLAUDE.md anyway, but the points you don't close now are the ones that cost most later.
Better to nail them down before you start coding.
CLAUDE.md you're about to generate has everything needed to launch an agent with this hierarchy: you can copy it and start.
When two levels conflict, the higher level wins, unless you've logged an explicit decision below.
Not set.
Not set.
Not set.
An agent's profitability isn't measured by token cost. It's measured by the Total Cost of Ownership (TCO) over 6 months — and it's systematically underestimated.
The cost of an AI agent in production isn't just what you pay the model provider. It splits across 3 layers, and only one shows up on the pricing page. The other two you discover along the way.
The visible part: what Claude, GPT, or your provider charges for input and output tokens.
The only layer that shows up in pricing.
Monitoring, vector database, cloud compute, logs, message queues.
$150–360/month minimum.
Reviewing the agent's responses, debugging when it drifts, updating context.
Tends to dominate the cost when human supervision is intense.
| Agent type | C1 Inference | C2 Orchestration | C3 Maintenance | 6-month total |
|---|---|---|---|---|
| Simple agent | $600 | $1,200 | $2,700 | $4,500 |
| Multi-tool | $6,000 | $5,400 | $13,200 | $24,600 |
| Multi-agent | $18,000 | $15,000 | $30,000 | $63,000 |
Pick a preset scenario or tune the values to your case. Figures update instantly. If you don't know what to put in a field, leave the preset: these are expert-judgment orders of magnitude, meant for you to edit. In particular, the infrastructure multiplier and the maintenance hours have no published figure — that's why they're inputs, not constants.
Sensitivity band: (optimistic) · (central) · (pessimistic)
This is what the agent will actually cost you. If most of it sits in C3, the bottleneck is human, not tokens. The band varies the infrastructure multiplier and the maintenance hours by ±50% — they are editable expert judgment, not published data — and the freed-hour recovery rate between 0.2 and 0.8. The central value is the headline number; the band exists to kill false precision.
This is the share of agent runs that finish cleanly end-to-end, after several months in operation.
We assume the freed hour is worth 40% of loaded cost — the midpoint of a 0.2–0.8 band (the rest is fixed cost that keeps running).
Band: (optimistic) · (central) · (pessimistic) h/month.
If the agent doesn't save you at least this many human hours each month, it doesn't pay for itself.
It's not enough for the agent to be cheap or fast. Before you delegate a task, answer these four questions. If even one comes back red, don't delegate yet: the cost of an error outweighs the savings.
Switching from Claude to GPT, or between models from the same provider, isn't free: you rewrite prompts, recalibrate edge cases, redo evals. An abstraction layer (a wrapper that isolates your code from the specific provider) has an up-front cost, but it sharply cuts the cost of every subsequent switch. This section tells you how many switches it takes before the layer pays off.
You pay the full migration cost every time you switch models.
You invest once in the layer and the following switches come out much cheaper.
It's the sum of everything you need to spend, directly or indirectly, to keep an agent running reliably in production over several months. It includes tokens, infrastructure, and human maintenance hours. It's not the same as the model price.
When an agent saves you an hour, that hour isn't worth the person's full salary: the rest (office, benefits, management) is fixed cost that keeps running even if the task disappears. The 0.4 we use is the midpoint of a 0.2–0.8 sensitivity band — no source pins down that exact value; the closest evidence (Bick, Blandin and Deming, NBER w32966) suggests the real capture of saved time is smaller than the mechanical saving. The results band in section 2 uses those extremes.
When an agent chains steps, the successes multiply. The probability the whole chain finishes cleanly is p to the N. With 95% per step and 20 steps: 0.95²⁰ = 0.36. Raising per-step reliability helps less than it seems: at 99% per step, 20 steps reach 82%, no more. And p^N is actually optimistic: per-step error isn't constant — when the model's context contains its own previous errors, it errs more often ("self-conditioning", Sinha et al., arXiv 2509.09677). That's why long chains need intermediate validation.
An agent doesn't stay equally good forever: the prompt ages, data changes, the provider updates the model underneath you. We model this degradation as p(T) = p0 · e^(-λT), where lambda tells you how much reliability you lose each month. A value of 0.02 means roughly 2% less reliability every month.
In tokens, Anthropic measured that an agent consumes ≈4× the tokens of a chat conversation, and a multi-agent system ≈15× the chat baseline (that is, ≈3.75× a single agent) (Anthropic, 2025). The infrastructure multiplier this calculator uses (×1 / ×4 / ×10) is editable expert judgment, with no published source: each extra agent adds monitoring, queues, error handling, and debugging of the interactions between them.
If you operate in the European Union, you have to add quality management, annual maintenance, and per-system validation. The overhead is real, but there is no reliable published figure that quantifies it: include it as its own line in your budget.
Switching from one model to another means rewriting prompts, recalibrating edge cases, and redoing evals. An abstraction layer (a wrapper that isolates your code from the specific provider) cuts that cost sharply past a certain number of switches. Section 4 calculates that point.
Sources: Anthropic (2025, multi-agent token consumption), Kanwat (2025) and Sinha et al. (2025, compound reliability), Pan et al. (2025, human supervision load), Bick/Blandin/Deming (NBER, value of the freed hour), Gartner (2025) and MIT NANDA (2025) (failure rates — predictions with caveats), DORA (2024 and 2025), METR (2025 and 2026 update), official pricing pages. The infrastructure multipliers, maintenance hours, and temporal drift are editable expert judgment — that's why they're inputs.
You haven't completed yet. The course is designed to be followed in order — each module builds on the previous one.
You can skip ahead if you want, but the material will be harder without the prior foundation.
The two tracks complement each other; Tools are transversal and useful any time.
You finished the track. Now you know how to use the AI that's already changing work.
) ═══════════════════════════════════════════════════════ -->