Hey there! Welcome to Platform Weekly. Your weekly tune-up of platform engineering engines. Every week, we pop the hood on what the community’s been building and send you the good stuff.
Plus...
Most platform teams we talk to sit somewhere between “we have an IDP” and “agents are shipping through it.” Very few know where they stand compared to everyone else. Fill out this survey to get your benchmarking score. As a thank-you, we’ll send you a free eBook version of Thinking in Platforms.
How do you run Kubernetes at scale without piling on more tools? Join Nutanix on November 19 for a live demo of Day-2 Kubernetes done right, with a production-ready platform!
You already know how to build an AI software factory
You’ve heard me tell the Ford story before. In 1908, building a Model T took about 12 hours. By 1913, it took about 1.5. Ford didn’t find faster workers. He redesigned the system they worked in, then kept treating that system as a product to improve. (*cough go read my book Thinking in Platforms *cough)
That idea is the whole premise of platform engineering.
But there’s one detail in that story I always skip. And right now, it might be the most useful bit.
The factory was always the point
Here’s what’s funny about the AI software factory hype. Platform engineering has been building factories for a decade. We just called them Internal Developer Platforms, and the people walking the golden paths were humans.
What changed is the users. Once agents start walking those paths, your IDP becomes an Agentic Engineering Platform (AEP), built for agents as well as people.
Give that AEP enough autonomy, and it becomes an AI software factory. Agents pick up the work, build it, validate it, and ship it. Humans stop doing the work and start orchestrating it, and eventually they just set the constraints. (That’s Level 3 and 4 in the four levels of agentic software development, if you want the full ladder.)
So no, you don’t need a brand new discipline. You need the one you already have, pointed at our favorite new user - the agent.
When does the value start?
In our last factory issue, I told you Level 3 is where the magic happens. Here’s the part I didn’t dig into.
In the State of AI in Platform Engineering Vol 2, we saw that 1 in 3 organizations have 2-3x throughput since adopting AI. No surprises there. The surprise is how that numbers changes depending on what level of agentic software development the org is.
At level 1, 30% of teams have 2-3x’d throughput. 36% at level 2. And an incredible 60% at Level 3.
Simply putting agents into the value stream (L2) adds just six points. Rebuilding the platform underneath them buys twenty-four. If you’ve read anything from us in the last few months, we’ve hammered you with that.
Why the jump? Because by Level 3, teams have built the guardrails. Governed paths across code, review and test. Every agent with its own identity, running in a sandbox, inside a spend limit. That’s what gives teams the confidence to actually let agents loose, and that’s where the real value shows up.
At Level 2, the agents are just as capable. The org isn’t. Without those mechanisms, letting agents run free is a massive risk engineers are genuinely afraid to take. Nobody wants to let it rip with an agent when there’s nothing in place to stop it getting them fired. So the agents stay on a short leash, and the gains stay small. (Self-reported numbers, ofc, but the shape is hard to ignore.)
At the same time, teams sprinkle agents across every stage of the SDLC at once. Agentic code gen here, an AI reviewer there, a test bot somewhere else. Loads of incomplete motion held back by platforms that aren’t ready. And not one line that runs from start to finish!
Build it like your first platform
Remember how the best platform teams built their first IDP? They didn’t boil the ocean. They picked one painful, frequent path, built it properly, got real users on it, and earned the right to build the next one.
Same playbook here.
Pick one boring class of work. High volume, low novelty. Dependency bumps, config fixes, flaky test repairs.
Give that one line everything. All five ingredients. Signals that kick off the work, a path with a clear definition of done, gates good enough to replace review for that class of change, identity and budget for every agent run, and context the agent can reason over.
Run it like a product. Who are its users? What does done look like? How much of the work ships without a human touching it, and what does it cost per accepted change?
Then, and only then, open line two.
Now for the detail I always skip;) In 1909, Ford announced he’d only build one car from then on. The Model T, with the exact same chassis every time. The factory only had to get one thing right, at huge volume, and the whole line could be built around it.
Find your Model T. Then build your first line around it.
Go see the lines that actually run
Want to see how teams built their factory lines all the way through? Go to PortCon, the AI Software Factory Summit. It’s November 17 and 18 at SHACK15 in the Ferry Building in San Francisco, with engineering leaders from NVIDIA, GitHub, SpaceX, Uber and Expedia sharing what worked and what didn’t. That includes GitHub’s Travis Gosselin and NVIDIA’s Aaron Erickson.
We’re in the era of the AI software factory, and the beauty of that is, what’s got to be done is no surprise for us platform folks, so get it.
Pick your Model T.
And as always... stay crunchy 🥐
Quick bites
Highlight of the week
What Ford did to crack the code for cars built the foundation for platform engineering. And it’s one of three main case studies in my book Thinking in Platforms, next to SIXT, and together they show how to run recurring work as a production system that both your engineers and your agents can rely on. Get your copy here! Or fill out the maturity survey to get a free eBook version.
Upcoming webinar
If we’re in the age of the AI software factory, you need to see one running before they can build their own. Port’s Field CTO Matar Peles shows how to go from governed agents to features that ship on their own, on December 10, 19:00 CET. Save your spot here. It’s the last webinar of 2026, so don’t miss out!
From the community
Missed a webinar? Catch up on the Platform Engineering YouTube channel.