Platform engineering in the agentic era · Working concept · October 2026 · Max Körbächer

The Infrastructure Factory

Everyone is building software factories. Nobody has built the thing they run on.

A platform that turns intents from humans or agents into compliant, owned, observable, cost-attributed environments, with no ticket in the path.

The shift

Code got cheap. The environments it needs still move through tickets.

Code supply75%of new code at Google is AI-generated and approved by engineers.Sundar Pichai, Google Cloud Next, April 20264
Delivery side+441%median time a change spends in review, while throughput per developer rose 33.7%.Faros AI, April 20267
Delivery side−7%main-branch throughput for the median team, while daily pipeline runs rose 59%.CircleCI, February 20268

Agents now write most new code at Google, Uber and Anthropic, and the pull requests arrive in minutes.465 What each one needs next, an environment with an owner, a network position and a budget, still moves through a ticket that a person opens a few times a day. Review time quadrupled and main-branch throughput fell. The gap between code supply and delivery capacity is where the queue forms.

All three figures are reported by the organization that benefits from them. Read them as direction, not precision.

Live model

The queue is the thesis.

Two lanes receive the same work at the same rate. The software lane ships at agent speed. The environment lane waits for a ticket.

Pull requests shipped0
Waiting for an environment0
Environments delivered0
Both lanes receive work at the same rate. The software lane ships at agent speed. The environment lane stops at a gate a person opens a few times a day. See the full model, with the policy-check switch, and the evidence.

Three roles

Infrastructure plays three roles in a software factory

It is under the loop, inside the loop, and the source of the next signal.

Infrastructure sits under the factory, flows through it, and feeds it alerts, incidents and logs become new signals Software factory loop Signals Plan Build Verify Release Monitor Product line: features and bug fixes Infrastructure line: Terraform PRs, incident remediation, CVE fixes Production runtime, clusters, cloud resources deploys runs on Factory floor: infrastructure the factory runs on Control plane dispatch, scheduling, budgets, audit trail Context and data plane MCP gateway, context lake, memory, session data Dev environments, IaC reproducible, previewable, agent-ready Compute and sandboxes isolated execution, VMs for computer use Models and inference routing, providers, self-hosted open models Cross-cutting identity, access policy, RBAC, audit observability and replay security and supply chain Highlighted: infrastructure in its three roles, the floor, a factory line, and production as the signal source.
D1. Three roles of infrastructure in a software factory: the floor it runs on, a line it runs, and production as the source of the next signal. Download PNG

The software factory sources agree on the first role and take the other two for granted. The floor is the infrastructure the loop runs on. The line is the infrastructure work the loop carries, which Port already runs as factory lines.44 The signal source is production itself: a factory that cannot see its environments cannot start the next loop. One platform supplies all three, and that platform is the infrastructure factory.

The floor it runs on

Control plane, context plane, environments, sandboxes, models and inference, with identity, policy, observability and supply chain across all of them.

A line it runs

Terraform pull requests, incident remediation and CVE fixes move through the same gateway, evaluations and release control as product code.

The signal source

Alerts, incidents and logs from production start the next loop. A factory without the floor has nothing to monitor.

What a software factory is, from its own sources

Not NVIDIA's AI factory

Not NVIDIA's AI factory, but it overlaps

NVIDIA's AI factory turns power into tokens.48 That is infrastructure specialised for accelerated compute, and under that reading it is an infrastructure factory with an unusual product mix.

The infrastructure factory produces environments: places to run things that arrive owned, compliant, observable and attributed to a cost center.

An AI factory in the model sense produces models fit for purpose: sourced, adapted, evaluated, registered and served.

The two share a floor and nothing else: GPU quotas, inference endpoints, registries, identity, policy, cost, observability and sandboxes. Every shared capability has exactly one owner.

The infrastructure and AI factories share a floor, not a product Infrastructure factory compliant places to run things · tenancy, regions, network zones · clusters, namespaces, environments · approval chains, cost centers · on-prem, sovereign, hyperscaler Quality signal lead time to a compliant environment AI factory, model factory models fit for purpose · sourcing and selection · fine-tuning and adaptation · evaluation and safety · lifecycle, drift, retirement Quality signal eval scores, drift, cost per token Shared floor capabilities both depend on · GPU capacity and quotas · inference endpoints, gateway · artifact and model registries · identity, policy, cost · observability, provenance · agent runtime, sandboxes Quality signal utilization, cost per tenant Platform engineering builds the shared floor. AI engineering builds the agents that run the factories.
D3. The infrastructure factory and the AI factory share a floor, not a product. Every capability in the middle has exactly one owner. Download PNG

How the three factories fit together

What it takes

What it takes to be one

  1. Intent in, environment out. The request names purpose, owner, data and lifetime. The product is a place to run things, not a list of resources.
  2. No ticket in the path. Humans decide once, at the contract. Nobody decides per request.
  3. Policy before provisioning. Allow or deny, with a reason the requester can act on, before anything is built.
  4. Ownership and expiry by default. Cost center, compliance owner and on-call are inherited. Everything has a lifetime.
  5. The factory measures itself. Lead time, zero-touch rate, time to owner and orphan rate, published to its customers.

Most teams are somewhere on this ladder. Find your stage, then read what the next one removes from the path.

  1. 0Ticket desk
  2. 1Automated fulfilment
  3. 2Self-service portal
  4. 3Contracted API
  5. 4Infrastructure factory

The twelve factors · The full ladder

Where to start

Start with the sandboxes

The first vending line is ephemeral sandboxes for coding agents. The demand exists already: sandboxes on Google Kubernetes Engine grew 16x in under five months after the Agent Sandbox preview.9 A sandbox needs an identity, a lifetime and a policy, and nothing a human has to sign. The second line is preview environments per pull request, where ownership defaults and cost attribution get exercised at volume. The third is production vending, by which time the contracts, policies and inventory already exist.

The three lines, in order