Founding cohort

Anyone can demo an agent. Very few can ship one.

The gap is not the model. It is everything around it: a persona that holds when someone tries to talk it out of the rules, a permission model that survives contact with real data, and an audit trail you would be happy to hand a client. That is what this course is about.

Join the list See the curriculum

Seats are not on sale yet. Nothing to pay, and no spam.

Six parts

  1. I Foundations: the loop, the identity split, the tool contract
  2. II The persona: six layers, and why backstory is a guardrail
  3. III Safety that holds: default-deny, provenance, triple enforcement
  4. IV Autonomy: schedules, context, and engineering silence
  5. V Knowledge and channels: retrieval without a vector database
  6. VI Operator track: governance, ROI and what to automate first

The gap

The tutorial ends where the hard part begins

Almost every agent tutorial stops at a working tool call. Then you point the thing at a real inbox, real customer records and real money, and you discover the questions nobody covered.

Someone emails your agent claiming to be the founder. What stops it believing them?

The agent can read the customer table. What stops that data appearing in an outbound reply?

Your API keys live in the same database the agent can query. Now what?

It runs unattended at 3am and wants to do something irreversible. Who approves it?

It has nothing useful to say this run. Does it stay quiet, or fill the channel with noise?

It got something wrong in front of a customer. Can you reconstruct exactly why?

Every one of these has a concrete, transferable answer. This course is those answers, in the order you need them.

Curriculum

Six parts, built in the order you will hit them

Written lessons and hands-on labs. Parts I to V are in both tiers. Part VI is the operator track.

6 parts
26 lessons
6 hands-on labs
I

Foundations

Both tiers

What an agent actually is once you strip the marketing off it.
  1. 01 The loop: prompt, model, tool call, result, repeat, and where it goes wrong
  2. 02 The identity split: whose credentials the agent uses versus whose permissions bound it
  3. 03 The tool contract: schemas, a normalised result vocabulary, one executor behind every transport
II

The persona

Both tiers

A six-layer architecture for writing an agent that stays itself under pressure.
  1. 01 WHO, WHY, HOW THEY THINK, HOW THEY ACT, HOW THEY INTERACT, SYSTEM
  2. 02 Backstory as guardrail anchor: give a rule a reason the model will not argue with
  3. 03 Fencing a dangerous trait at the moment you grant it
  4. 04 No trait is a lever: the clause that stops personality being quoted as permission
  5. 05 Registers and containment: which voice, which channel, and which way the failure defaults
III

Safety that holds

Both tiers

The part that separates a demo from something you let near a customer.
  1. 01 Default-deny, and four tiers of authorisation people actually use
  2. 02 Provenance over topic: gate where a fact came from, not what may be discussed
  3. 03 Counter-guardrails: stopping your safety rules from eating normal business
  4. 04 Triple enforcement: prompt, task, and code, with the code trusting none of the prompt
  5. 05 Secrets: why reading one is already the breach
  6. 06 Anti-fabrication gates: placeholders, unapproved numbers, invented specifics
IV

Autonomy

Both tiers

Agents that run on a schedule with nobody watching.
  1. 01 Cadence and context: how much history to carry, and why fresh beats cumulative
  2. 02 Engineering silence: making "say nothing" the default output
  3. 03 Confirm-first handshakes, and the propose-only override when no human is present
  4. 04 Caps, kill-switches and dry-run defaults that ship in the off position
V

Knowledge and channels

Both tiers

Giving an agent things to know and places to speak.
  1. 01 Pointer skills: a retrieval procedure instead of a documentation dump
  2. 02 When you genuinely need embeddings, and the many times you do not
  3. 03 Channel design: inbox, chat and social, with deduplication that is structural not hopeful
  4. 04 Result caps and paging, and why oversized tool results kill agents
VI

Operator track

Operator

For the person who wants the system, not the source code.
  1. 01 Mapping your business processes onto an allow-list
  2. 02 Choosing the first thing to automate, and the arithmetic behind that choice
  3. 03 Governance: review cadence, audit, and what to do the day it gets something wrong
  4. 04 Live cohort session, plus a written review of one persona you bring

Every part ends with a lab you run yourself. Nothing here is watch-only.

What you actually take away

Eight techniques you will not find in the docs

These came out of running an agent against a live business, not out of a whiteboard. Each one exists because something went wrong without it.

01

Backstory as guardrail anchor

Give the agent a costly past failure whose lesson is exactly the rule you need. Rules argued from character do not get argued away.

02

Fence the trait as you grant it

Any capability that could be misused gets its boundary written into the same sentence that introduces it.

03

No trait is a lever

One clause that makes personality unusable as authorisation. It closes the most common jailbreak against a rich persona.

04

Capability is not permission

Describe what the agent can do and what it may do in separate sections that reference each other. Conflating them is how agents overreach.

05

Provenance over topic

Gate information by where it came from, not by subject. This is what lets an open, chatty agent stay safe with confidential data.

06

Named counter-guardrails

Explicit clauses that stop your safety rules over-firing. Without them a well-guarded agent refuses to do its job.

07

Triple enforcement

Every real constraint appears in the prompt, in the task, and in code, with the code treating the prompt as untrusted.

08

Pointer skills

A short retrieval procedure that stands in for an entire documentation corpus, with one editable source of truth.

Who it is for

Two people get the most out of this

The developer

You can already call an API and wire up a tool. You want the architecture and the failure modes, so the thing you build survives a client putting real data through it.

  • Comfortable in Python and a terminal
  • Has shipped something that talks to an API
  • No prior agent or machine learning experience needed

The technical founder or operator

You want an agent running parts of your business, and you need to decide what it is allowed to touch, in what order, and how you would know if it went wrong.

  • Can read code, does not have to write it
  • Owns the decision about what gets automated
  • Wants the governance answers, not just the build

Founding cohort

Seats open soon

The curriculum is set and the material is being written. Leave your email and you get first refusal at the founding price.

Builder

For the person writing the code.

$249 $349

one payment, lifetime access

  • Parts I to V, every lesson
  • Hands-on labs and the starter repository
  • The eight techniques, written up in full
  • Every future update to the material
Tell me when seats open

Operator

For the person who owns the outcome.

$749 $999

one payment, lifetime access

  • Everything in Builder
  • Part VI, the operator track
  • The persona boilerplate and review checklist
  • A live session with the cohort and the instructor
  • A written review of one persona you bring
Tell me when seats open

No payment, no card, one email when seats open. Unsubscribe in a click.

What you are buying, plainly

This is a pre-sale. The curriculum is fixed, the material is being written, and none of it exists for you to read today. Here is exactly what that means.

You get

  • Everything in your tier, delivered in one go on the start date
  • Founding pricing, locked, including future updates
  • A say in what gets covered, we will ask

You do not get

  • Anything to read before the start date
  • The source code of our own agent, this is a course, not a product licence
  • A certificate, an accreditation, or a job guarantee

FAQ

Questions, answered

What happens on the start date?

Everything in your tier unlocks at once: the written lessons, the labs and the starter repository. Operator seats also get the live session scheduled and the persona review opened. You will get an email with your access link.

What if the start date slips?

You get told before it happens, with a reason and a new date, and your refund window reopens for two weeks from that email. You are never stuck waiting on a promise with no way out.

Which model or framework does it teach?

None specifically, and that is deliberate. Everything here is about the architecture around the model: personas, permissions, tool contracts, autonomy and audit. It transfers to whichever provider and framework you use, and it will still apply after they change.

Do I need to know Odoo?

No. The examples are generic and the labs are self-contained. Odoo developers will recognise a lot of it, but nothing depends on knowing the platform.

Is this based on a real system?

Yes. Every technique in Part III came out of running an agent against a live business inbox, a real Stripe account and a public social presence. The course is a case study of that, not a distribution of its source code.

Builder or Operator, which one?

If you will be writing the agent, Builder. If you will be deciding what an agent is allowed to do in your business and answering for it, Operator. If you are doing both, Operator.

Can I upgrade later?

Yes. Email us and you pay the difference at founding prices, whenever you like.

How do refunds work?

Reply to your receipt within the refund window and we send the money back in full. No form, no reason needed, no follow-up email trying to talk you out of it.

Build the agent you would trust with your own business.

That is a much higher bar than a working demo, and it is a learnable one. The founding cohort is where it starts.

Join the list Talk to us

Free plan, no credit card. Your servers stay yours.