The Factory That Fits on a Laptop
Compliant by Design — Essay 03
This essay almost opened with a claim of invention. I had a phrase I liked, “the compliance-native software factory,” and a plan to plant a flag on it. Before putting it in print I spent an evening checking whether anyone had got there first. They had. Consulting firms, infrastructure vendors, the US Department of Defense. It was the most useful search I ran all month, because what it left me with is better than a flag.
The phrase is taken at the top
BCG Platinion published “The Agentic Software Factory” in March 2026. Their definition: autonomous AI agents build, test, and ship software around the clock while humans define business intent and review outcomes. The line that matters for this series is their compliance claim, quoted exactly: the agentic factory “does not make compliance harder, it makes it structurally easier,” because an intent-driven pipeline produces an audit trail of every decision as a by-product of working.
Pulumi followed in May with “Seven Rules for Building an AI-Native Software Factory.” It reads like field notes rather than strategy: pick tools your agents can drive, don’t let any agent mark its own homework, get your governance and risk people onside early or, in the author’s words, “AI becomes a compliance fight instead of a structural advantage.”
Behind both sits more prior art. 8090.ai sells AI-native development to regulated industries as its whole pitch. The US DoD has been doing compliance-as-code under continuous authority to operate for the better part of a decade. Nobody writing in 2026 invented this, and that includes me.
Rule seven
Pulumi’s seventh rule is the one I keep arguing with: “Run the factory in the cloud, not on a laptop.”
At their scale it’s right. Their reference build gives every client two dedicated AWS accounts, one production, one digital twin. If you’re running fleets of agents for enterprise clients, a laptop is a liability.
But I read that rule the week my factory shipped its second essay, opened its own pull request, passed its own compliance checks, and deployed, all from hardware that fits in a backpack. The enterprise version of the factory needs a cloud because it’s built for hundreds of engineers’ worth of output. The version I actually needed had to fit one founder, no platform team, and Australian regulation. That version fits on a laptop. Not because a laptop is the ideal substrate, but because the factory isn’t the hardware. It’s the gates.
What mine looks like
Concretely, the whole thing: a topic inbox in a database that captures incidents as they happen. Claude drafts from it in a voice I’ve corrected twice and now mostly don’t have to. I review for 30 to 45 minutes, factual accuracy and anything that doesn’t sound like me. Claude Code opens a pull request against a protected branch. CI runs gitleaks, so no credential reaches the repo, and a 16-assertion test suite that includes the site’s central claim as an executable check: zero script tags and zero third-party origins on the live pages. Nothing merges red. Merge deploys.
The gates weren’t designed as a system. Every one of them is scar tissue. The gitleaks stage exists because I committed a live API key to a public repo, which was essay one of this series. The banned-claims review exists because my marketing site wore a HIPAA badge for weeks, in a country where HIPAA means nothing, which was essay two. The pattern each time was the same: the incident happened, the fix became a check, and the check now runs whether or not I’m paying attention. Carefulness didn’t scale. The pipeline did.
That is, at one-person scale, exactly BCG’s claim about audit trails, except I didn’t build it to satisfy their thesis. I built it because I kept making mistakes and got tired of relying on my own vigilance.
The bottom is empty
Here’s what that evening of searching didn’t find. Everything published under this framing is enterprise: BCG writes for CIOs, Pulumi for platform teams, 8090 for institutional clients. I could not find the Australian version at the bottom of the market. Nobody is showing what compliance gates inside an AI-driven pipeline look like under the Privacy Act’s APPs, the Notifiable Data Breaches scheme, and the Essential Eight, at the scale of one founder or a five-person team, with working receipts.
The Australian commentary that does exist is about compliance of AI, mostly the automated-decision transparency requirements landing in December 2026. Important, and I’ll come back to it. But it’s a different subject. That’s regulating what AI does to people. This series is about what regulation does inside the pipeline that AI builds with. The first is well covered. The second, at my scale, in this jurisdiction, is an empty room.
I’m turning my own gates into an Australian compliance pack another founder could drop into a repo: the scanning config, the banned-claims list, the pre-ship sweep mapped to the APPs and the NDB scheme. It isn’t public yet. If that’s something you’d use, subscribe below and tell me, because what I build first depends on who’s asking.
The position I ended up with
The search was meant to protect a novelty claim. It killed the claim and handed me something sturdier: BCG says the agentic factory is the future of regulated software delivery. Fine. Here’s mine, running, on a laptop, under Australian rules, and the next essays in this series are the incidents it caught. Not a roadmap. A record.
“I invented this” is a claim about everyone else, and it dies the moment someone searches. “Mine is running” is a claim about you, and no consulting deck can take it away. If your positioning can’t survive twenty minutes of looking, better to spend the twenty minutes before print than after.
Compliant by Design is a fortnightly series on building AI products under Australian regulation — Privacy Act, APPs, NDB, health data — written from inside the work, not above it.