Why a platform
Building an Internal Developer Platform is a shared journey, for humans and agents
Por que uma plataforma
Construir uma Internal Developer Platform é uma jornada compartilhada, para humanos e agentes
What we want to say: building an IDP (Internal Developer Platform) is not deploying the final drawing. It is a journey we take together, with clear goals.
When we started, the idea was to show the drawing of the five planes. On the way it became clear it involves much more: the catalog and ownership, contracts and gates, identity, cost, security, and agents as new users of the platform.
Route of the series: this deck is the problem; the next ones are where we are heading, each plane piece by piece, the API and its clients, the journey from the idea to a running service, the full drawing, and what changes when the user is an agent.
Goal: agree on what the platform is and what it is not, and leave the room with allies.
O que queremos dizer: construir um IDP (Internal Developer Platform) não é implantar o desenho final. É uma jornada que fazemos juntos, com objetivos claros.
Quando começamos, a ideia era mostrar o desenho dos cinco planes. No caminho ficou claro que ele envolve muito mais: o catálogo e o ownership, contratos e gates, identidade, custo, segurança, e agentes como novos usuários da plataforma.
Roteiro da série: este deck é o problema; os próximos são para onde vamos, cada plane peça por peça, a API e os seus clientes, a jornada da ideia a um serviço rodando, o desenho completo, e o que muda quando o usuário é um agente.
Objetivo: alinhar o que a plataforma é e o que não é, e sair daqui com aliados.
Who absorbs the complexity today?
Quem absorve a complexidade hoje?
Infrastructure teams look after the infrastructure, and whoever has an idea still has to ask another team
Asking for infrastructure today
- Request: to another team
- Knowledge: documented, not self-service
Times de infraestrutura cuidam da infra, e quem tem uma ideia ainda precisa pedir a outro time
Pedir infraestrutura hoje
- Pedido: a outro time
- Conhecimento: documentado, não self-service
What we want to say: the developer does not carry the infrastructure alone, an infrastructure team looks after it. What still weighs is the path: whoever has an idea has to ask, and gets the package that exists.
Request: even when creating an environment is already automated for whoever operates it (a pull request handled by an operator, for example), it is still a request to another team, which has its own schedule.
Knowledge: the "how to" is written down, in how-tos and runbooks, and part of it already runs by itself. But it was written for whoever operates: whoever has the idea does not run it alone. The platform turns that same knowledge into a golden path the developer uses without asking, and that an agent uses too, because an agent only uses what is written in a format a machine reads (DORA, DevOps Research and Assessment, ROI report, p. 41).
Ask the room for an example of their own: it works well here.
Source: references/dora-roi-ai.md
Sources:
O que queremos dizer: o dev não carrega a infra sozinho, um time de infraestrutura cuida dela. O que ainda pesa é o caminho: quem tem uma ideia precisa pedir, e recebe o pacote que existe.
Pedido: mesmo quando a criação de um ambiente já é automatizada para quem opera (um pull request conduzido por um operator, por exemplo), continua sendo um pedido a outro time, que tem a sua própria agenda.
Conhecimento: o "como se faz" está escrito, em how-tos e runbooks, e parte dele já roda sozinho. Mas foi escrito para quem opera: quem tem a ideia não o executa por conta própria. A plataforma leva esse mesmo conhecimento a um golden path que o dev usa sem pedir, e que um agente também usa, porque um agente só usa o que está escrito num formato que a máquina lê (DORA, DevOps Research and Assessment, relatório de ROI, p. 41).
Peça à plateia um exemplo próprio: funciona bem aqui.
Fonte: references/dora-roi-ai.md
Fontes:
Today people absorb the complexity one ticket at a time; a platform absorbs it once, as self-service
Who absorbs the complexity
- Today: infra teams, one ticket at a time
- Platform: once, as self-service
Hoje pessoas absorvem a complexidade um ticket por vez; a plataforma a absorve uma vez, como self-service
Quem absorve a complexidade
- Hoje: times de infra, um ticket por vez
- Plataforma: uma vez, como self-service
What we want to say: at the starting point the developer does not carry the whole stack, the infrastructure teams do, one request at a time. A platform absorbs that complexity once, and the request becomes self-service.
The complexity of cloud native does not go away: either people absorb it on every request, or the people who build the platform absorb it once. The first model grows with the number of infrastructure people; the second does not.
DORA ROI (return on investment, p. 40): "a platform that treats the developer—or any builder—as the user, minimizing the cognitive load required to move code from an agent's prompt to a production environment."
David Iyanu Jonathan, "The platform engineering illusion" (platformengineering.com, 2026-07-31): "you build it, you run it" became "you build it, you run it, you secure it, you monitor it, you cost-optimize it, and you keep current on eleven CVE feeds." The platform's job is to "reduce unnecessary complexity while preserving necessary complexity", without hiding so much that debugging becomes archaeology.
Source: references/dora-roi-ai.md, references/platform-illusion.md
Sources:
O que queremos dizer: no ponto de partida o dev não carrega a stack inteira, os times de infraestrutura carregam, um pedido por vez. A plataforma absorve essa complexidade uma vez, e o pedido vira self-service.
A complexidade do cloud native não some: ou pessoas a absorvem a cada pedido, ou quem constrói a plataforma a absorve uma vez. O primeiro modelo cresce com o número de pessoas de infraestrutura; o segundo, não.
DORA ROI (retorno sobre investimento, p. 40): "a platform that treats the developer—or any builder—as the user, minimizing the cognitive load required to move code from an agent's prompt to a production environment."
David Iyanu Jonathan, "The platform engineering illusion" (platformengineering.com, 2026-07-31): o "you build it, you run it" virou "you build it, you run it, you secure it, you monitor it, you cost-optimize it, and you keep current on eleven CVE feeds." O trabalho da plataforma é "reduce unnecessary complexity while preserving necessary complexity", sem esconder tanto que depurar vire arqueologia.
Fonte: references/dora-roi-ai.md, references/platform-illusion.md
Fontes:
A powerful platform that is hard to use is not adopted: what is missing is empathy, not an API
The empathy gap
- Symptom: powerful but hard to use
- Question: how does it feel to use?
Plataforma poderosa e difícil de usar não é adotada: falta empatia, não API
A falta de empatia
- Sintoma: poderosa, mas difícil de usar
- Pergunta: como é usar?
What we want to say: a powerful platform that is hard to use is not adopted; what is missing is empathy, not an API.
Tatiana Mikhaleva, "The empathy gap: Why your platform needs UX, not just APIs" (platformengineering.org, 2026-03-23): IDPs fail by investing in the infrastructure and forgetting who uses it.
Two quotes from the article: "We forget the most critical component in the stack. We forget the human being trying to use it." And: "A backend engineer should not need deep Kubernetes knowledge to get a URL for their service."
Caveat: an opinion piece; the figures in the text have no original source, do not use them.
A second voice, the same thesis, in "The platform engineering illusion": "the platform was correct and nobody wanted it." The team built the platform the way you build a bridge, and developers lived it "the way you experience a toll booth". Also opinion: the figures of the account stay off the screen.
Source: references/empathy-gap.md, references/platform-illusion.md
Sources:
O que queremos dizer: plataforma poderosa e difícil de usar não é adotada; falta empatia, não API.
Tatiana Mikhaleva, "The empathy gap: Why your platform needs UX, not just APIs" (platformengineering.org, 2026-03-23): IDPs falham por investir na infraestrutura e esquecer quem a usa.
Duas citações do artigo: "We forget the most critical component in the stack. We forget the human being trying to use it." E: "A backend engineer should not need deep Kubernetes knowledge to get a URL for their service."
Ressalva: artigo de opinião; os números citados no texto não têm fonte original, não usar.
Uma segunda voz, a mesma tese, em "The platform engineering illusion": "the platform was correct and nobody wanted it." O time construiu a plataforma como quem constrói uma ponte, e o dev a viveu "the way you experience a toll booth". Também opinião: os números do relato ficam fora da tela.
Fonte: references/empathy-gap.md, references/platform-illusion.md
Fontes:
Why has AI not paid off yet?
Por que a IA ainda não deu retorno?
AI sped up writing code, but the return has not come yet for most
Since adopting AI, organizations report
- Throughput: 38% ship 2x or more
- Return: only 8% call it transformative
Source: State of AI in Platform Engineering, Vol. 2 (2026), pp. 20-22
A IA acelerou a escrita de código, mas o retorno ainda não veio para a maioria
Desde que adotaram IA, as organizações relatam
- Throughput: 38% entregam 2x ou mais
- Retorno: só 8% o chamam de transformador
Fonte: State of AI in Platform Engineering, Vol. 2 (2026), pp. 20-22
What we want to say: AI raised the volume delivered, but almost nobody saw the return in the business.
These are two different questions of the survey (State of AI in Platform Engineering Vol. 2, Weave Intelligence, 2026).
Throughput: how much more engineering delivers since adopting AI (Figure 7, p. 20). The 38% add those who reported 2x to 3x (29%) and 5x or more (9%).
Return: the impact of AI on the whole life cycle (Figure 9, p. 22). Only 8% call it transformative; 29% are still experimenting, with no value realized, and 14% have a negative return once licences, tokens and review are counted.
The message is the report's own: "Throughput is not return" (p. 22). Shipping more features is not the same as shipping more business results. That gap is the thread of the whole report (p. 8).
Caveat: self-reported data, 350 respondents.
Source: references/state-of-ai-in-platform-engineering.md
Sources:
O que queremos dizer: a IA aumentou o volume entregue, mas quase ninguém viu o retorno no negócio.
São duas perguntas diferentes da pesquisa (State of AI in Platform Engineering Vol. 2, Weave Intelligence, 2026).
Throughput: quanto a engenharia passou a entregar desde que adotou IA (Figure 7, p. 20). Os 38% somam quem relatou 2x a 3x (29%) e 5x ou mais (9%).
Retorno: o impacto da IA no ciclo de vida inteiro (Figure 9, p. 22). Só 8% o descrevem como transformador; 29% ainda estão em experimentação, sem valor realizado, e 14% têm retorno negativo quando se contam licenças, tokens e revisão.
A mensagem é a do próprio relatório: "Throughput is not return" (p. 22). Entregar mais features não é o mesmo que entregar mais resultado de negócio. Essa lacuna é o fio do relatório inteiro (p. 8).
Ressalva: dado autorrelatado, 350 respondentes.
Fonte: references/state-of-ai-in-platform-engineering.md
Fontes:
The obstacle is not the model, it is the platform around it
Main blockers to scaling AI
- #1 Platform readiness: cited by 27%
- #4 Model capability: cited by 11%
Source: State of AI in Platform Engineering, Vol. 2 (2026), p. 24
O obstáculo não é o modelo, é a plataforma ao redor dele
Principais obstáculos para escalar a IA
- #1 Prontidão da plataforma: citada por 27%
- #4 Capacidade do modelo: citada por 11%
Fonte: State of AI in Platform Engineering, Vol. 2 (2026), p. 24
What we want to say: what holds AI back is the platform around the model, not the model.
State of AI Vol. 2, p. 24: "platform readiness" leads the blockers to scaling AI (27%); "model capability" comes fourth (11%).
A supporting quote: "It's because they have the platform around the model." (Rickey Zachary, Thoughtworks, p. 24). Nobody is waiting for a better model.
Source: references/state-of-ai-in-platform-engineering.md
Sources:
O que queremos dizer: o que trava a IA é a plataforma em volta do modelo, não o modelo.
State of AI Vol. 2, p. 24: "platform readiness" lidera os obstáculos para escalar IA (27%); "model capability" fica em quarto (11%).
Citação de apoio: "It's because they have the platform around the model." (Rickey Zachary, Thoughtworks, p. 24). Ninguém está esperando um modelo melhor.
Fonte: references/state-of-ai-in-platform-engineering.md
Fontes:
Agents without rebuilding the platform yield little; rebuilding the platform yields four times more
Share that at least doubled throughput
- Agents added (Level 1 → 2): 30% → 36%
- Platform rebuilt (Level 2 → 3): 36% → 60%
Source: State of AI in Platform Engineering, Vol. 2 (2026), pp. 18, 21
Agentes sem reconstruir a plataforma rendem pouco; reconstruir a plataforma rende quatro vezes mais
Parcela que ao menos dobrou o throughput
- Agentes adicionados (Level 1 → 2): 30% → 36%
- Plataforma reconstruída (Level 2 → 3): 36% → 60%
Fonte: State of AI in Platform Engineering, Vol. 2 (2026), pp. 18, 21
What we want to say: putting agents on the same platform yields little; the jump comes from rebuilding the platform.
The same throughput question as the previous slide, now cut by level of agentic maturity (Figure 8, p. 21): the share of each level that at least doubled delivery since adopting AI.
The levels measure how much work the agent carries (Figure 5, p. 18): at Level 1 the AI suggests and a person executes; at Level 2 agents open pull requests in parallel, and a person triggers and checks; at Level 3 agents work continuously from observed signals, and review becomes the exception.
Level 1 to 2: agents arrive, the platform stays the same, +6 p.p. (percentage points). Level 2 to 3: the platform is rebuilt, with governed paths a machine can run in code, review and test, +24 p.p. Maturity pays in steps, not in a gradient.
Caveats, said aloud and not on screen: the levels and gains are self-reported (p. 18, p. 21); the Level 3 sample is small, n = 50 (Level 1: 143; Level 2: 110; Level 4: only 14); Syntasso, the company behind Kratix, is a partner of the report. It is a direction, not a precise measurement.
Source: references/state-of-ai-in-platform-engineering.md
Sources:
O que queremos dizer: pôr agentes sobre a mesma plataforma rende pouco; o salto vem de reconstruir a plataforma.
A mesma pergunta de throughput do slide anterior, agora cortada por nível de maturidade agêntica (Figure 8, p. 21): a parcela de cada nível que pelo menos dobrou a entrega desde que adotou IA.
Os níveis medem quanto trabalho o agente carrega (Figure 5, p. 18): no Level 1 a IA sugere e uma pessoa executa; no Level 2 agentes abrem pull requests em paralelo, e uma pessoa dispara e verifica; no Level 3 agentes trabalham continuamente a partir de sinais observados, e o review vira exceção.
Level 1 para 2: entram os agentes, a plataforma continua igual, +6 p.p. (pontos percentuais). Level 2 para 3: a plataforma é reconstruída, com paths governados e executáveis por máquina em código, review e teste, +24 p.p. A maturidade paga em degrau, não em gradiente.
Ressalvas, ditas em voz alta e não na tela: os níveis e os ganhos são autorrelatados (p. 18, p. 21); a amostra do Level 3 é pequena, n = 50 (Level 1: 143; Level 2: 110; Level 4: só 14); a Syntasso, empresa do Kratix, é parceira do relatório. É direção, não medição precisa.
Fonte: references/state-of-ai-in-platform-engineering.md
Fontes:
AI amplifies the system that already exists: with a weak platform, the effect disappears
AI amplifies what exists
- Weak platform: no effect
- Quality platform: strong effect
Source: DORA State of AI-assisted Software Development (2025), p. 71
A IA amplifica o sistema que já existe: com plataforma fraca, o efeito some
A IA amplifica o que existe
- Plataforma fraca: sem efeito
- Plataforma de qualidade: efeito forte
Fonte: DORA State of AI-assisted Software Development (2025), p. 71
What we want to say: AI amplifies the system that already exists; without a quality platform, the effect disappears.
DORA 2025, p. 71: "AI adoption has a negligible effect on organizational performance when platform quality is low, but when platform quality is high, the effect is strong and positive." DORA is DevOps Research and Assessment.
It is a neutral source (Google and DORA, 4,867 respondents) that confirms the previous slide without Syntasso's bias.
A concrete picture from DORA ROI (p. 13): code generated faster "simply piles up in front of manual security reviews or brittle deployment pipelines".
Caveat: a quality platform also brings a little more instability and friction (DORA 2025, p. 62 and p. 70).
A counterpoint, in "The platform engineering illusion": AI does not solve adoption, because "adoption was never a tooling problem to begin with — it was a trust and design problem". The tool that generates the config got smarter; the developer's trust does not move faster because of it.
Source: references/dora-ai-report.md, references/dora-roi-ai.md, references/platform-illusion.md
Sources:
- State of AI-assisted Software Development 2025
- The ROI of AI-assisted Software Development
- The Platform Engineering Illusion
O que queremos dizer: a IA amplifica o sistema que já existe; sem plataforma de qualidade, o efeito some.
DORA 2025, p. 71: "AI adoption has a negligible effect on organizational performance when platform quality is low, but when platform quality is high, the effect is strong and positive." DORA é DevOps Research and Assessment.
É uma fonte neutra (Google e DORA, 4.867 respondentes) que confirma o slide anterior sem o viés da Syntasso.
Uma imagem concreta do DORA ROI (p. 13): o código gerado mais rápido "simply piles up in front of manual security reviews or brittle deployment pipelines".
Ressalva: plataforma de qualidade também traz um pouco mais de instabilidade e fricção (DORA 2025, p. 62 e p. 70).
Contraponto, em "The platform engineering illusion": a IA não resolve adoção, porque "adoption was never a tooling problem to begin with — it was a trust and design problem". A ferramenta que gera a config ficou mais esperta; a confiança do dev não anda mais rápido por isso.
Fonte: references/dora-ai-report.md, references/dora-roi-ai.md, references/platform-illusion.md
Fontes:
Who is the platform for?
Para quem é a plataforma?
Treat the platform as a product: the developer is a customer, not someone who opens tickets
Developers are the customers
- Self-service, not ticket queues
- Enablement, not enforcement
Tratar a plataforma como produto: o dev é cliente, não quem abre ticket
O dev é o cliente
- Self-service, não fila de tickets
- Habilitar, não impor
What we want to say: the platform is a product, and the developer is its customer.
Back to the first slide of the act: tickets and waits become self-service, and the "how to" that was tribal knowledge becomes a product. The report sums it up: "The model is based on enablement rather than enforcement and self-service rather than ticket queues." (Platform as a Product, p. 3).
A product has users, a value proposition, a feedback loop and a focus on adoption (pp. 3-4, 9-11). It is not an infrastructure project with a delivery date: it has a roadmap and is measured by those who use it.
Caveat: the four authors work for Syntasso, the company behind Kratix.
In the same line, in "The platform engineering illusion": "A technology platform is infrastructure with an API. A developer platform is a product with users", with preferences, workarounds and a memory of every time the tool let them down. And: "every manual step in a "self-service" system is really just a support queue wearing a costume."
Source: references/platform-as-a-product.md, references/platform-illusion.md
Sources:
O que queremos dizer: a plataforma é um produto, e o dev é o cliente dela.
Volta ao primeiro slide do ato: tickets e esperas viram self-service, e o "como se faz" que era conhecimento tribal vira produto. O relatório resume: "The model is based on enablement rather than enforcement and self-service rather than ticket queues." (Platform as a Product, p. 3).
Produto tem usuários, proposta de valor, ciclo de feedback e foco em adoção (pp. 3-4, 9-11). Não é projeto de infraestrutura com data de entrega: tem roadmap e é medido por quem usa.
Ressalva: os quatro autores trabalham na Syntasso, a empresa por trás do Kratix.
Na mesma linha, em "The platform engineering illusion": "A technology platform is infrastructure with an API. A developer platform is a product with users", com preferências, workarounds e memória de cada vez que a ferramenta os deixou mal. E: "every manual step in a "self-service" system is really just a support queue wearing a costume."
Fonte: references/platform-as-a-product.md, references/platform-illusion.md
Fontes:
Sources
Fontes
- The ROI of AI-assisted Software Development · DORA and Google Cloud · 2026
- The Platform Engineering Illusion · David Iyanu Jonathan · 2026
- The empathy gap: Why your platform needs UX, not just APIs · Tatiana Mikhaleva · 2026
- State of AI in Platform Engineering, Volume 2 · Sam Barlien, Luca Galante and Florian Lipp (Weave Intelligence) · 2026
- State of AI-assisted Software Development 2025 · DORA and Google Cloud · 2025
- Platform as a Product · Abby Bangser, Daniel Bryant, Colin Humphreys and Cat Morris · 2025
The ROI of AI-assisted Software Development
DORA and Google Cloud · 2026
- About: A framework to compute the return on AI in software delivery, built on the findings of the DORA 2025 report, with a calculator of example assumptions.
- For this talk: It names the platform's role in the agentic era: "a platform that treats the developer—or any builder—as the user, minimizing the cognitive load required to move code from an agent's prompt to a production environment" (p. 40).
The ROI of AI-assisted Software Development
DORA and Google Cloud · 2026
- Sobre: Um framework para calcular o retorno da IA na entrega de software, apoiado nos achados do relatório DORA 2025, com uma calculadora de premissas de exemplo.
- Para esta palestra: Nomeia o papel da plataforma na era agêntica: "a platform that treats the developer—or any builder—as the user, minimizing the cognitive load required to move code from an agent's prompt to a production environment" (p. 40).
The Platform Engineering Illusion
David Iyanu Jonathan · 2026
- About: An opinion piece: internal developer platforms fail because they are run as tooling initiatives when they are a product discipline.
- For this talk: "A technology platform is infrastructure with an API. A developer platform is a product with users." Its figures are the author's own account and stay off the screen.
The Platform Engineering Illusion
David Iyanu Jonathan · 2026
- Sobre: Um artigo de opinião: as plataformas internas falham porque são tocadas como iniciativas de ferramentas quando são uma disciplina de produto.
- Para esta palestra: "A technology platform is infrastructure with an API. A developer platform is a product with users." Os números são o relato do próprio autor e ficam fora da tela.
The empathy gap: Why your platform needs UX, not just APIs
Tatiana Mikhaleva · 2026
- About: An opinion piece: platforms invest in infrastructure and forget the person who uses it, and adoption stays low.
- For this talk: "A backend engineer should not need deep Kubernetes knowledge to get a URL for their service."
The empathy gap: Why your platform needs UX, not just APIs
Tatiana Mikhaleva · 2026
- Sobre: Um artigo de opinião: as plataformas investem na infraestrutura e esquecem quem a usa, e a adoção fica baixa.
- Para esta palestra: "A backend engineer should not need deep Kubernetes knowledge to get a URL for their service."
https://platformengineering.org/blog/the-empathy-gap-why-your-platform-needs-ux-not-just-apis
State of AI in Platform Engineering, Volume 2
Sam Barlien, Luca Galante and Florian Lipp (Weave Intelligence) · 2026
- About: A survey of 350 practitioners, April to August 2026, on how organizations adopt AI in platform engineering; maturity levels and gains are self-reported.
- For this talk: "Throughput is not return" (p. 22): 38% ship at least twice as much since adopting AI, only 8% call the return transformative, and platform readiness is the first blocker (p. 24).
State of AI in Platform Engineering, Volume 2
Sam Barlien, Luca Galante and Florian Lipp (Weave Intelligence) · 2026
- Sobre: Uma pesquisa com 350 profissionais, de abril a agosto de 2026, sobre como as organizações adotam IA em platform engineering; os níveis de maturidade e os ganhos são autorrelatados.
- Para esta palestra: "Throughput is not return" (p. 22): 38% entregam pelo menos o dobro desde que adotaram IA, só 8% chamam o retorno de transformador, e a prontidão da plataforma é o primeiro obstáculo (p. 24).
https://weaveintelligence.io/research/state-of-ai-in-platform-engineering-2026
State of AI-assisted Software Development 2025
DORA and Google Cloud · 2025
- About: DORA's 2025 report, a survey of 4,867 technology professionals on AI in software delivery.
- For this talk: "AI adoption has a negligible effect on organizational performance when platform quality is low, but when platform quality is high, the effect is strong and positive" (p. 71).
State of AI-assisted Software Development 2025
DORA and Google Cloud · 2025
- Sobre: O relatório DORA de 2025, uma pesquisa com 4.867 profissionais de tecnologia sobre IA na entrega de software.
- Para esta palestra: "AI adoption has a negligible effect on organizational performance when platform quality is low, but when platform quality is high, the effect is strong and positive" (p. 71).
Platform as a Product
Abby Bangser, Daniel Bryant, Colin Humphreys and Cat Morris · 2025
- About: An O'Reilly report: a platform is a product, with users, a value proposition, a feedback loop and a focus on adoption, not an infrastructure project with a delivery date.
- For this talk: "The model is based on enablement rather than enforcement and self-service rather than ticket queues" (p. 3). The four authors work for Syntasso, the company behind Kratix.
Platform as a Product
Abby Bangser, Daniel Bryant, Colin Humphreys and Cat Morris · 2025
- Sobre: Um relatório da O'Reilly: plataforma é produto, com usuários, proposta de valor, ciclo de feedback e foco em adoção, não um projeto de infraestrutura com data de entrega.
- Para esta palestra: "The model is based on enablement rather than enforcement and self-service rather than ticket queues" (p. 3). Os quatro autores trabalham na Syntasso, a empresa por trás do Kratix.
https://www.syntasso.io/platform-as-a-product-oreilly-report