Brownfield factories, promises, and the demand factory

From Rube Goldberg loops to mode discernment

McCarthy opens with a pun on Diffusion—diffusing model capability out into the hundred-trillion-dollar global economy—and tells the room this talk is dual-channel: human framing on stage, plus a closing QR link (diffusion.io/brownfield) packed with technical verification papers for your agents. Back in July 2025 on Claude 4, getting factory loops to stabilize required enormous token spend and fragile hand-wiring:

“You know, even then, we could see glimmers that this thing was going to work. But at tremendous token expense, and only in the presence of a real. Rube Goldberg action.”00:00:52

Now that software factories have moved into the conference mainstream, the mistake is applying one process to every engineering mode:

“Begin to use some discernment about whether we're discovering product market fit, we're breaking new ground. We're conquering new territory, or whether we are doing something that, you know, already has a precedent and is already working.”00:02:12

Evidence: transcript 00:00:00, 00:00:52, 00:02:12

Why greenfield stalls and brownfield wins

When building something brand new, teams rarely know what they want—like a commuter asking for a faster horse in a Word spec. Trying to industrialize ambiguous discovery only manufactures the wrong software faster:

“So that's of course the time for me to say, please stop building software factories.”00:02:12

Brownfield codebases flip the problem on its head. Customers already vote with dollars, daily utilization, and support tickets—giving your loops an existing behavioral oracle so teams can clear technical debt and rewrite legacy bottlenecks without guessing what to build:

“Is it's something that we're already being paid for. It's something where customers are saying with their dollars and their and their utilization. Like yeah, this works for me. I like it. I want more of this.”00:03:18

“What I'm using is what I want. Okay, so we know what to build. We build we build the existing thing.”00:03:18

Evidence: transcript 00:02:12, 00:03:18

Promises as invariants in continuous loops

Diffusion calls each core product guarantee—a feature, cost ceiling, price, or security boundary—a promise: an invariant you refuse to regress on. Encoding promises into closed loops only stabilizes a system when runtime truth feeds back; cut the feedback leg, and an amplifier just amplifies noise:

“Feature is a promise. You know the cost and price of a thing is a promise.”00:04:52

“As many of you know, this is the sort of idealized block diagram amplifier, and if you cut that feedback leg. You're just going to amplify noise, okay?”00:04:52

With promises encoded, commits stop following human daily and weekly seasonality and run around the clock—constantly checking whether anything got slower or less secure:

“So another way to summarize that is, you know, my brownfield factory on an existing set of promises, it should just be. Continuous. It should be background. It should be keeping those promises.”00:06:30

Evidence: transcript 00:04:52, 00:06:30

Factories simplify over time

Builders who struggled with sprawling agent scaffolding a year ago are watching harnesses generalize and skills shrink—with most of the credit going to stronger underlying models:

“The skills are getting shorter. Harnesses are getting more general, of course, and a lot of this again is credit to the models themselves.”00:08:09

He maps that curve onto SpaceX’s Raptor 1, Raptor 2, and Raptor 3 rocket engines. Raptor 1 flew despite a maze of external piping; each successor packed higher thrust into fewer moving parts. In aerospace, that took armies of rocket engineers—whereas software teams inherit simplification automatically every model release:

“We have been allies in the simplification of work because we have Astro, Table, and Carbon. Right, and so they are trying to take our factory from version one to two to three.”00:08:09

Evidence: transcript 00:08:09

One week to faster, cheaper, safer

Give a brownfield factory a clean instruction—the same system, without trading safety for speed along the Pareto curve—and leave subjective “better” for human taste and discovery:

“So, so it should be faster. It should be cheaper. It should be safer.”00:10:27

“So I think everyone should have the ground field factory. You should have it for your existing commitments. You should have it. It should become better, faster, and cheaper over time. I find that this takes like a week.”00:10:27

The companion library behind his closing QR code (diffusion.io/brownfield) lays out the verification primitives behind safe rewrites: shadow replay and differential testing, coverage-guided fuzzing, property-based and deterministic simulation, mutation testing and held-out checks, specification proofs (including AWS Cedar), simulated users, and canary cutover.

Evidence: transcript 00:10:27, 00:17:07

Elevate your ambition: the demand factory

Once a brownfield factory keeps promises running reliably, the real constraint flips. Teams are churning out working software—only to realize buyers weren’t asking for more buttons to click:

“They actually, they actually weren't asking for software. They were asking for more gross margin. They were asking for more customers. They were asking for more brand differentiation, okay, and just giving them more windows to click on, and buttons to click on. That's not doing it for them anymore.”00:12:03

Because engineers fluent with models can build supply at will, they inherit a new job: building a Demand Factory to rate-match market learning to software output and expand the company’s brand envelope—like Spindrift launching an energy drink (credible adjacent promise) rather than motor oil:

“So that means like we have a new job now. We have to figure out how to build a demand factory. Because our brownfield factory can't build the moat. Right. All it can do is keep our existing promises running well.”00:13:17

“And then figure out what is the maximum extent of what I can possibly promise to the market, and that's my brand envelope.”00:15:32

“Maybe you can't sell motor oil, but maybe you can sell an energy drink. So figure out what your energy drink is and start selling it now.”00:15:32

Evidence: transcript 00:12:03, 00:13:17, 00:15:32, 00:16:12

Lower-latency demand and the first-mover edge

Even in a world where every competitor rents the same frontier models, the team that pairs fast agents with sharper market questioning, overnight synthesis, and rapid elasticity experiments wins a latency edge—causing customers to stop self-censoring and ask for more:

“I think your demand factory helps you ask better questions of your market. Because the agents can help prioritize what to ask next.”00:13:17

Expanding across customers, wallet share, SKUs, and coherent product lines eventually runs out of organic headroom—leaving consolidation to whoever built the faster loop first:

“Okay, and if you get started now, you're going to be the one doing the acquiring.”00:16:12

• Reserve factories for brownfield promises: stop industrializing unvalidated greenfield guesses; use continuous background loops where paying usage already defines invariants (00:02:12, 00:03:18, 00:04:52).
• Rely on model-driven simplification: like Raptor 1→2→3, cleaner models slim down skills and harnesses while verification gates keep rewrites faster, cheaper, and safer (00:08:09, 00:10:27).
• Pair supply with a Demand Factory: use agents to ask better market questions, allocate experiments, and sell adjacent lines inside your brand envelope (00:13:17, 00:15:32, 00:16:12).

Evidence: transcript 00:13:17, 00:16:12, 00:17:07