Building the Factory

We Built a New Factory Around Software's 1993 Moment

In 1993, Porsche built its last handmade car, the 911 version 993. The software industry just had our 1993 moment — and we built a new factory around it.

For 60 years, every Porsche that left Stuttgart was built by hand. Engine by hand. Body panels beaten by hand. Wheels fitted by hand. It was an extraordinary craft and it was slowly killing the company.

Toyota had spent two decades systematizing manufacturing without sacrificing quality. When Porsche hired their consultants in the early '90s, the CEO walked the factory floor on day one, chainsaw in hand, and cut the 12-foot shelving units in half. "Human scale," he said. They weren't bolting lean principles onto their existing process. They were rebuilding the factory around new assumptions entirely.

What followed wasn't a reduction in quality. Porsche became one of the most profitable companies in the world making the most high quality and high performance cars on the planet. The craft got better.

The Inflection Point

Most organizations are bolting AI onto an existing process.

Agentic AI can now write production-quality code, review its own output, and iterate on feedback. That's not a productivity feature. It's a structural shift in what software development is.

The common response is predictable: layer AI tools onto existing processes. Small efficiency wins with an easy to measure experience. Frontier labs — Anthropic, OpenAI, Google DeepMind, xAI — aren't building better tools. They're replacing the knowledge worker production model itself.

Software is strategic because it's the only extreme knowledge task that's also fully verifiable in an automated way. But more importantly, software is the substrate of every other knowledge industry. Automate the ability to build it and you unlock the ability to automate legal analysis, financial modeling, medical documentation, logistics planning — every domain that runs on information and rules. Software is not one industry among many. It is the master key to all of them.

"I don't write any code anymore. I let the model write it — I just edit."Dario Amodei, CEO of Anthropic

The frontier models themselves are a commodity. They will continue to improve. They will continue to compete on price and capability. Your data, your processes, your customer relationships, and your institutional knowledge — these are the raw materials no lab can replicate. But raw materials sitting in a warehouse are not a business. They become a business only when you build a factory around them.

A few organizations will do something fundamentally different — redesign the entire lifecycle around the assumption that agents are first-class participants in the work. We chose to be one of them.

What We Built

Seven stages. Two phases. One deliberate gate.

Discovery (Stages 1–4) exists to surface so much context, about the client's industry, their competitive landscape, their actual workflows, therefore, what gets handed to execution is nearly impossible to build wrong. Deep industry and competitive research that previously required a large consultancy? We do that now, with agents, at a higher quality level. We can spin up multiple agent-driven explorations of an idea, put working software in front of a client in days, and feed real feedback forward. The closer to the reality of production features we get early in the cycle, the better and less risky software becomes. It's a new superpower for us and our clients.

The Commitment Gate is the formal point where we say: we know what to build, we know why, and we're committing to scope and architecture before a single line of production code is written. It prevents the two failure modes we see most often — locking in too early before the problem is understood, and staying in exploration so long that delivery suffers. The gate isn't bureaucracy. It's permission to experiment freely before it, and discipline to execute cleanly after it.

In Execution, agents handle well-specified work under human review. Engineers lead on architectural and security-critical decisions. Nothing ships without human accountability. PRs don't go away. If an agent wrote it, an engineer owns it. That standard doesn't soften because the author was a machine. They increase because we can review more of the system at scale for more quality check points.

The Part Most Organizations Miss

The SDLC isn't a line. It's a cycle.

Every deployment feeds real production behavior back into the working context. Every cycle makes the system smarter — not just the individuals on the project, but the scaffold, the harnesses, the factory itself across Phase2_.

We're building skills files that make agents smarter on the next cycle than they were on the last. Prompts, context packages, and harnesses that encode what we've learned and carry it forward across every engagement. This is what Toyota figured out that everyone else only partially copied. When every person in the organization has responsibility for improving the system, the system compounds.

Our measure of success isn't just what we ship for a client. It's whether the factory we use to build it is smarter than the one we started with six months ago.

What This Asks of the Team

More, not less.

The honest version of this change is that it asks more of everyone — at a higher level of abstraction, with more leverage, on harder problems.

Engineers: Less time writing code directly. More time engineering systems, reviewing agent output, and making architectural decisions that now scale faster and carry more consequence.

Designers: Shaping agent interaction patterns — a discipline that barely existed two years ago — alongside traditional UX and increasingly robust prototyping work.

Client Leads: Evaluating working prototypes within hours of ideation instead of weeks. Reading deep client context surfaced by research agents and improving it continuously.

Project Managers: Orchestrating a faster, more iterative cycle with different rhythms. The loop is tighter. The decisions come sooner. The feedback is more real.

The fear we hear most — that agents will replace judgment — misunderstands the problem. Judgment doesn't shrink. It scales. When an agent can implement a well-specified feature, a world-class engineer shouldn't be the one typing it out. That engineer should be deciding whether the feature is designed correctly, whether the architecture will scale, whether the security model holds. An engineering judgment that used to affect one feature now affects a million lines of code and every future cycle of the factory.

Version 1.0. By Design.

Recursive improvement is the product.

This system is imperfect at launch. That's intentional. The goal isn't a perfect process on day one. The goal is a factory that gets smarter than the one we started with. We'll tear it apart at Summer Summit. Rebuild what needs rebuilding. Ship a better version to ourselves for the second half of the year.

That's what Porsche learned in 1993. Not that manufacturing was the enemy of craft. That the right kind of manufacturing was craft — elevated to a scale no individual could achieve alone.

We intend to build that.

‍