The software factory of the near future

01Everything was hand-rolled.

One factory line. Five phases. Each discipline owns a stage and stays in its lane. Work moves by hand-off — and stalls when people leave for meetings and docs that never touch the repo. By the time build starts, the PRD is already stale.

02AI shortened the build.

Over the last year AI walked into the build phase, and most teams bolted it onto the factory line they already had. That’s Level 1: AI in the IDE, everyone trying a different tool, token bill climbing. You’d expect the whole factory line to shrink. Nobody fully trusts what the machine wrote, so review grows and the Test belt gets longer. The feature arrives… barely sooner.

87% say AI helps them ship faster; 28% feel less friction. — Atlassian, State of DevEx 2026

03Then came vibe coding.

In 2025 Andrej Karpathy named it: English is the programming language. Open an AI app, describe the idea, often get working software back in one shot. Look under the chat window and it’s a factory of one — one person, one belt, on a board that’s all AI.

Karpathy coined the term in February 2025. — Vibe coding, Wikipedia

04A small group — or potentially one person — runs the whole factory line.

It doesn’t always cover the non-functional requirements a mature engineering team would. Don’t underplay it anyway. It’s changing how people express ideas and win buy-in — and a small group, or potentially one person, can run the lot.

05The CRAFT stack underneath.

Under that factory of one sits CRAFT: Context, Reasoning, Authoring, Functions, Toolchain. Most layers swap. Context — intent, specs, code, domain — is yours. Build the loom. Don’t weave the cloth.

The models are becoming a commodity; the context is what’s yours. — Collibra

06CRAFT lets one person run a SPIRAL.

Scout → Pinpoint → Invent → Review → Activate → Learn. Agents help at each step. You still own the judgment.

07Demos, not memos.

Level 2 takes what works in vibe coding into the enterprise. The factory of one drops into the AI-augmented factory line where discovery and design used to live, and becomes the Demo. A small group, or potentially one person, can run it, so it’s much shorter. By hand, the team writes a PRD into a folder and nothing works until the end. Here, a few working prototypes get built, used, and compared. Keep what sticks. Sell it with a demo — demos, not memos.

Matt Waite coined the phrase in 2009 while building PolitiFact: demos make the abstract concrete; memos keep everything easy on paper. — Matt Waite, “Demos, not memos”

08Through the gate, the prototype becomes a spec.

Once the team picks the demo that works best, it passes a gate that turns it into a specification — just in time. That’s the intent the machines will build from.

09Spend the time on the spec.

Flattened out, the spec is a feature file: intent, scenarios as given / when / then, acceptance criteria, plus the designs and context behind it. This is where the time goes now. A precise spec keeps the rest of the factory line running. Teams can write skills that check every spec is CLEAR: clear, lean, executable, accurate, responsible.

Feature files use Gherkin, the syntax behind Cucumber. — Cucumber, Gherkin reference

10A second gate turns the spec into product.

The spec rides the factory line to a coding agent that builds shippable product, story by story. A story ships only when its evals pass. If one fails, a reviewing agent stops the factory line before the defect spreads. Who merges? Treat PRs like one-way and two-way doors: AI sorts them by risk, people review the risky ones, and each organisation sets its own bar.

11Spec, build, validate, release.

So the AI-augmented factory line runs from the demo and the spec — they agree what to build and why. One person plans and builds, another works the spec, and they validate together until everything passes. When a story fails, the gate turns red and the factory line stops itself — Toyota’s Andon (Japanese for “lantern”: the light that signals a problem so someone comes to help). By hand, that same defect would ride all the way to the repo. With the factory line AI-augmented, it’s much shorter than doing things by hand, only a little longer than vibe coding — and it ships safely.

Andon lights up when a factory line stops so someone can help. — Toyota Production System

12Release through the gates.

The aim hasn’t changed: get every new feature idea in front of customers as soon as possible, through a series of gates — internal colleagues, an experiment, a public beta, then general availability. Push as many ideas through as you can. Not every one will make it. That’s fine.

13Smaller groups run whole factory lines.

A factory line that once needed a whole team can now be run by a small group, so the same team can run several factory lines at once. Pull back and the factory line copies itself — each one operated by a few people, release gates fanning into rings. That’s where the velocity comes from. Senior talent can coach others as an expert guild: a separate group off to the side that teams turn to for questions and support.

14We enter the era of disposable software.

The gates are now confidence rings around the factory. Most boxes stop at an inner ring — and that’s the system working. Throwing most of what we build away would’ve made us uncomfortable a couple of years ago.

15How an AI-augmented factory is set up.

From above it’s a plan: the team in the middle, a factory line for each person or small group, rings around them. Like any agile team, it runs on feedback loops. An idea that doesn’t make the gates comes back for refinement. Every return makes the next lap faster.

16The foundations.

Lift the floor and the team stands on a company brain, shared CI/CD, the models it runs on, and the company’s own CRAFT stack. Beside it sits CRAFT Ops — keep the stack’s health metrics green, keep optimising. Above them sits the expert guild.

17Agents run the loops.

So far, people still run most of the loops. In Level 3, every station on every factory line gets an agent: one shapes the demo, one drafts the spec, one writes the code, one tests it against the acceptance criteria and evals. They talk along the factory line and across factory lines. People own both ends — intent at the start, judgement at the end. Because the factory lines watch themselves, one person can watch many, and step in only when a factory line can’t fix itself. When a factory line stops, the team and CRAFT Ops fix the cause, not just the code: a new rule or skill goes back into the CRAFT stack. Its health metrics become integral.

18Agents use computers too.

Not every task is code. Agents can use computers and apps — clicking, typing, filing through tools like Playwright. Remember the rooms that pulled people off the factory line? Now an agent takes the meeting notes and files the actions, triages the inbox and drafts the replies, builds the update deck from the data. The latest models even make motion graphics for product marketing.

19The flywheel.

Go top-down and there’s a flywheel. The team, CRAFT Ops and the expert guild become cogs, with agents under every part of the engine. The faster it spins, the more good ideas make it out through the rings.

20Three levels of AI adoption.

Level 1: AI coding. Build gets faster; the rest is by hand — tool sprawl, token burn, many people per factory line.

Level 2: demos, not memos. Prototype first, CLEAR specs, release gates — fewer people per factory line, still running every loop.

Level 3: agentic loops and a defined tool set. Agents run the loops on top of CRAFT, CRAFT Ops and good evals. One or two people can run several factory lines.

Where’s your team?

21A factory of factories.

Zoom out and every team runs its own factory on one shared AI-augmented platform — a factory of factories. Anyone can send an idea into another team’s factory as a pull request; that team reviews it in time they’ve set aside. With that many ideas moving, the pipeline needs managing.

49% spend 4+ hours a week reviewing AI output (Atlassian, State of DevEx 2026). 64.9% of reviewers report lower initial trust in AI-authored PRs (Rahman et al., Plausible Is Not Enough).

22Plan in CHIPS, not estimates.

Estimation as we knew it is dead, in my view — when code for a feature takes on average an hour to write, we need to focus on measuring the time required to get alignment on a buildable spec instead. Same for issue tracking: why keep story tickets when the spec lives with the code in the repo? Teams plan in CHIPS instead: Choose a key result; Hypothesise solutions that will turn the dial; Identify the blockers to eliminate; Players and a time box; Score the outcome against the key result and the original intent.

23Plan to remove blockers.

Quarterly metrics rarely change. Line up the bets and clear blockers early. Some bets fail — that’s fine. Keep time for BAU and risky reviews. Let AI self-drive KTLO. Yearly strategy sets themes; the scoreboard shows which bets moved the needle.

24Most ideas die cheaply. The best ones ship faster.

That’s my view of the software factory of the near future: AI at every station, people running several factory lines at once, and clear ways to automate the factory lines — and to stop them. For those of us already in the current SDLC / PDLC, I think this is enormous opportunity — work smarter, get customer value out faster, and have more fun doing it. Most ideas will die cheaply. The best will ship much faster than before.

I’d love to hear what you think. Follow me on X or Substack — I’ll share the skills I’m building to run this factory.