The seriesTen artifactsSheet 27

Specifications, not a product.

What the series actually gives you.

Three books, and between them ten artifacts specified precisely enough to implement: the shape a routing policy takes, what a capability contract must declare, what a delivery guardrail has to decide before an agent may act. None of it is a library you install. All of it is a shape you can build in the stack you already run.

§ 01The premise

A shape you can build, in your stack.

Most engineering books hand you either a philosophy you cannot act on or an implementation that assumes your infrastructure looks like the author's. Neither survives contact with a regulated institution, which already has a stack, already has constraints, and cannot adopt someone else's opinionated framework wholesale.

So the series specifies shapes rather than implementations. A schema states what a thing must declare and what must be true of it, and leaves how you store, serve and enforce it to you. That is deliberately less convenient than a library and considerably more portable, because a shape can be honoured by a Kubernetes CRD, a config service, or a table in a database you already run.

§ 02The artifacts

Ten things you can actually build.

Book
1

LLM Systems in Production

The infrastructure layer. Out 24 August 2026.

RoutingPolicy CRD reference
The declarative object that decides which model may answer a given request, at what latency, cost and risk tier. Promoted through the same review path as any other change, so a routing decision is versioned rather than configured by hand.
Routing-event log schema
What a routing decision must record to be reconstructable afterwards. This is the difference between knowing a model answered and being able to show why that model was chosen for that request.
Book
2

Prompt Systems & Agent Orchestration

The application layer. Out 5 September 2026.

Capability contract
What a single agent is permitted to do, declared rather than implied. It is the object that turns least privilege from an intention into something the runtime can enforce at the moment of a tool call.
Agent graph schema
How multiple agents compose into one accountable workflow, with the boundaries the flow may not cross stated in the object rather than inferred from the code that runs it.
Critique object
The structured form a verifier's judgment takes, so that a rejection carries a reason a downstream module and a human reviewer can both act on. See the PEVG pattern for where it sits.
Reasoning Profile
The dials that tune how a planner decomposes work by business line: step budget, exploration appetite, retry budget, and how strict the verifier will be held downstream. One planner, dialled differently, becomes a careful auditor or a quick analyst.
Book
3

DevOps for AI-Native Platforms

The operations layer. Out 15 September 2026.

Delivery Guardrail schema
Which actions an agent may execute alone and which must escalate, declared in advance by policy rather than decided in the moment by the agent that wants to act.
Risk profile
The tiering object that determines how much autonomy a given system earns, and therefore what it must clear before it ships.
PARA catalog entry
How an operational agent's role, allowed actions and success metrics are recorded so it can be governed. A role and a reach the catalog cannot state are an agent nobody can govern. See the PARA operating model.
Golden-path template
The paved route a team follows to ship something governed by default, so that the compliant path is also the easiest one. Governance that is harder than the workaround is governance that will be worked around.

Each book additionally carries a compliance crosswalk aligning its material to TOGAF, DMBOK, the EU AI Act and the NIST AI Risk Management Framework, which is the artifact most readers reach for first when someone asks how the architecture maps to a standard the institution already reports against.

§ 03What it is not

What the series does not give you.

It is not a codebase

There is no repository to clone and no package to install. Specifying a shape rather than shipping an implementation is a deliberate trade: it costs you the convenience of working code and buys portability into a stack that was never going to match the author's anyway. If what you need is a library, this is the wrong purchase.

The institution in it is fictional, and stated to be

The systems are presented through a deliberately fictional regulated bank so the method can be shown end to end without exposing a client. Nebula Financial is a teaching device, not a case study being passed off as a deployment. The patterns behind the artifacts are implemented in Maxim, which is live, and in client engagements which are confidential. Where a claim rests on a real system, it says so; where it rests on the fiction, it says that too.

All three books are still forthcoming

Book 1 is out 24 August 2026, Book 2 on 5 September, Book 3 on 15 September. This page describes what they specify. It is not a substitute for them, and nothing here should be read as the schemas themselves.

§ 04Questions

What readers ask.

Does the series include working code?

No, and it does not claim to. What it specifies are schemas, contracts and templates: the shape a routing policy takes, what a capability contract must declare, what a delivery guardrail has to decide. Those are implementable in whatever stack you run, which is the point of specifying a shape rather than shipping an implementation that assumes your stack looks like someone else's.

Are these artifacts drawn from a real deployment?

They are presented through a deliberately fictional regulated bank so the full method can be shown end to end without exposing a client. The patterns behind them are what the author implements in Maxim, which is a live product, and in client engagements which are confidential. Nebula Financial is a device for teaching, not a case study being passed off as real.

Do I need all three books?

No. The books are layered rather than sequential: infrastructure, application, operations. A platform team standing up model routing needs the first, a team building agents needs the second, a team running them in production needs the third. The Cross-Book Navigation Guide on the series resources page exists to route you to the one that matches the problem you actually have.

Can I use these artifacts commercially?

The licensing of the series text and its free companion materials is set out with those materials rather than asserted here, and it differs between the paid book text and the free companions. If you intend to adapt a schema into a commercial product, ask and you will get a straight answer rather than a link to a licence page that does not quite address it.

When are the books available?

All three are forthcoming. LLM Systems in Production is out 24 August 2026, Prompt Systems and Agent Orchestration on 5 September 2026, and DevOps for AI-Native Platforms on 15 September 2026. The already-published Enterprise Playbook on AI governance for the Middle East is a separate book and available now.

§ 05Start

Start with the layer that is failing.

If you already know which layer is your problem, go straight to that book. If you do not, the free Readiness Self-Assessment will tell you which of the four governance layers is lowest, and that is usually the same place the engineering is thinnest.

Fin · Build
See the series →