Skip to main content

Platform as a Product

The agentic platform in this section is, underneath the new vocabulary, still a platform. The discipline Camille Fournier and Ian Nowland lay out in Platform Engineering: A Guide for Technical, Product, and People Leaders did not stop applying when the workloads became agents — if anything, it matters more. Before the context layer and the workflows, get the frame right.

Platform engineering is the discipline of developing and operating platforms. The goal of this discipline is to manage overall system complexity in order to deliver leverage to the business.

Platform Engineering, Fournier & Nowland — Ch. 1

That word — leverage — is the whole point. The work of a few engineers on a platform team reduces the work of the entire organization: it makes application engineers more productive and eliminates duplicated effort. An agentic platform earns its keep the same way, or it is just an expensive demo.

Product
not a project — it has customers and a roadmap
Curated
opinionated self-service, not a pile of primitives
Leverage
a few engineers reduce work for the whole org
Adoption
is engineering — earned, never mandated

Three lenses at once

The book's subtitle is its thesis: a platform is technical, product, and people work simultaneously. Get one wrong and the platform stalls — a beautifully engineered tool nobody adopts, or a beloved idea that falls over in production.

T

Technical

architecture · leverage

Abstract a curated set of primitives into a paved road that reduces cognitive load.

  • paved roads, not mazes
  • self-service infrastructure
  • high operational standards
P

Product

developers are customers

Treat internal engineers as users with a roadmap, not ticket-filers in a queue.

  • a roadmap and real users
  • adoption is the metric
  • a product, not a project
U

People

change management

The hardest half. Skeptics by default; you bring them along by demonstrating value.

  • skeptics by default
  • champions scale you
  • demonstrate, never mandate
The book is organized around what success looks like: platforms that are aligned, trusted, manage complexity, and are loved. None of those are purely technical.

A platform is a product, not a project

The single most important idea in the book is also the one teams most often skip. A platform has customers — your own engineers — a roadmap, and a success metric that is adoption, not lines shipped. Projects end; products are maintained, evolved, and held accountable to the people who depend on them. The moment you treat an internal platform as a one-time delivery, it starts to rot.

With the word “product” we strive to achieve for platforms what Steve Jobs created with Apple products: against a broad range of demand for features the product is deliberately and tastefully curated, both through what it does and, more importantly, through what it leaves out.

Platform Engineering, Fournier & Nowland — Ch. 1

Knowing your customer is not a metaphor here — it is the prerequisite. The developer persona is who every roadmap decision answers to.


Curated, not just self-service

The word the book keeps returning to is curated. A platform is self-service — you do not file a ticket and wait — but it is not a pile of primitives either. "The cloud" is not a platform; it is an overwhelming array of offerings. The platform is the opinionated, paved road built on top: the common path is one command with sane defaults, and the parts you could get wrong are already decided for you.

The tension between curation and flexibility has a name in the book — "building blocks, not batteries included." A batteries-included platform that wires everything together end to end becomes too coupled to evolve; foundational, composable blocks with escape hatches let trusted teams unblock themselves.

For internal engineering platforms the value of stability and future flexibility cannot be sacrificed in the name of ideal usability.

Platform Engineering, Fournier & Nowland — Ch. 12

A golden path that cannot be left becomes a cage, and engineers route around cages. The art is making the paved road so obviously the path of least resistance that almost nobody needs the hatch — while leaving it there for the cases that do.


Buy the commodity, build what is yours

The oldest platform decision is build versus buy, and the rule is the durable one: buy what is commodity, build what differentiates you. AI does not change the rule, only where the line falls. You rent the frontier model and the managed vector store; you do not rebuild your observability stack or your code host. But the grounding, the gates, and the golden paths encode how your team works — and nobody sells those.

Buy the commodity
Foundation models
rent the frontier, never train it
Vector DB / search
managed retrieval infrastructure
Observability backend
tracing, metrics, logs — already paid for
CI / code host
do not rebuild these
Build the differentiation
Context layer
your team’s knowledge; nobody sells this
Orchestration layer
how agents run your SDLC and gates
Eval suite
what “good” means in your domain
Golden paths
the opinionated road for your stack
The model is rented; the grounding, gates, and golden paths are not. Building the commodity is how platform teams become a bottleneck instead of a source of leverage.

Measure it like a product

Because a platform is a product, you measure it like one — and the vanity metrics (agents deployed, lines generated) tell you nothing. The metrics that matter are about the customers.

Sometimes platform engineers confuse the platform itself with the outcomes they’re trying to drive — the existence of a new platform isn’t an outcome.

Platform Engineering, Fournier & Nowland — Ch. 12
What an internal platform is measured oncustomer signals, not output counts
Adoptionplatform
Share of teams and engineers actively using it.
The only number that says the platform is real — and a lagging signal of trust, not a leading one.
Time to first valueplatform
Minutes from zero to a working golden path.
Onboarding cost, measured. Reduce it and adoption compounds on its own.
Reliability & trustplatform
Uptime, and the confidently-wrong rate.
Trust is lost in a single surprise; reliability is an adoption feature, not just an engineering one.
Would-be-missedplatform
Would engineers be upset if it vanished tomorrow.
The truest product signal an internal platform has. Ask it directly.
The book’s measurement stance is “manage it like a product” — satisfaction, adoption, leverage, operational SLOs — not a scorecard of platform-internal activity.

For the agent-quality side of measurement — VCR, rubric scores, cost per successful outcome — see Metrics That Matter.


NetflixWorkflow Orchestrationsource ↗
Maestro: declarative orchestration as a platform primitive

Netflix's Maestro workflow orchestrator is a study in the curated self-service principle. Rather than giving teams raw primitives (a scheduler, a queue, a retry library), the platform team built a declarative YAML DSL that handles scheduling, dependency resolution, retry semantics, and failure recovery as defaults. Teams express what they want — this step depends on that one, retry three times on transient error — without reasoning about how it's implemented. The abstraction is opinionated enough to eliminate the common cases, composable enough to handle the edge cases, and observable enough that when something goes wrong, the platform team can diagnose it without the product team's help. That last property — the platform can support itself — is what makes self-service scale.

Platform teams own the engine; product teams own the logic. The split eliminated a class of operational burden from every team that onboarded — without removing their ability to express complex dependency graphs.

Put trust ahead of adoption

You do not get adoption by decree. The book is blunt that "standardization via authority isn't enough" — you earn the move by making the new path good enough that the old one withers, and by sequencing rollout by risk tolerance: onboard less-critical use cases first, gather the operational data, and use that trust to win the next, more critical tranche.

1

Earn, don’t mandate

Demonstrate leverage on a real, painful task. A mandate produces malicious compliance and shadow tooling, not converts.

2

Sequence by risk

Less-critical apps first. There is no compression algorithm for experience — accelerate the curve, don’t skip it.

3

Shadow platforms are signal

When a team routes around you, sometimes they’re right. Let good ideas prove out, then merge the ones with real demand.

Adoption is a loop, not a launch — see Adoption & Enablement for the full play.

Field note

We launched with a full wizard UI: twelve screens, forty-two fields, a live preview. Engineers said it was beautiful. Then we watched them not use it. The wizard optimized for discoverability; engineers optimized for speed. They already knew what they wanted — they just needed a clean way to express it. We replaced the wizard with one YAML template and a make run. Adoption went from twelve teams to forty-one in two quarters. The lesson: time to first value is the adoption metric. Everything else is product-team self-congratulation.

internal platform launch, month three
The adoption test

Adoption is the platform's only honest metric. Everything else — features shipped, pipelines deployed, lines of generated code — is internal accounting. If engineers would not be upset if the platform disappeared tomorrow, it hasn't earned its place.

  • Mandate produces malicious compliance — engineers will use the platform exactly enough to say they did.
  • Shadow platforms are signal, not insubordination. A team that built their own version is telling you something about your onboarding cost.
  • The platform is loved when its absence would be felt. Ask directly: would you be upset if this disappeared tomorrow?

The one tension AI adds

There is a single tension an agentic platform adds to the classic list, and it is the one that matters most: a platform that ships code faster also ships bad code faster, unless quality is engineered in. The job is to push the speed-and-quality frontier outward — more throughput and fewer regressions. Three disciplines keep that honest — grounding so agents hallucinate less, evaluation so "it feels better" becomes a measured number, and accountability so every action stays reviewable and reversible. The Overview treats this tension as the spine of the whole section; the point here is only that it survives the shift to agents.

The model under the platform will keep changing, faster than any book can track. The discipline of building platforms — product mindset, curated self-service, build-versus-buy, adoption as engineering — does not. Learn it once and it transfers to whatever the workload becomes next.


Referenced: Camille Fournier & Ian Nowland, Platform Engineering: A Guide for Technical, Product, and People Leaders (O'Reilly, 2024).