Spec2Ship: the factory of factories

What it has built

• CarerSupport: the participant-owned NDIS care record. In production.
• Digital Starter Kit: small-business websites, live in days, flat fee. A factory built by the factory.
• MyWealthTracker: AI guidance grounded in Australian tax and super law. Pilot.
• ProductMatch: AI pricing intelligence for retailers. Live.
• Dealroom: two AI agents negotiating a deal for their humans. In the build queue.
• Spec2Ship itself: the factory’s most demanding customer.

Built to be handed over, not to depend on me. Every agent, rule and check lives in versioned, documented and tested configuration, like any other product. Every new production line is installed from that same source. It’s designed to be installed in a team, not to live in one person’s head.

Swipe sideways to see the whole diagram

The production line

Eight stages. Each one answers a question an executive should ask of any AI-built system.

Swipe sideways to see the whole diagram

1. Shape: “Do we know what we’re building?”

Spec before code

• Does: a crew of specialist agents takes an idea from rough to buildable before any code is written. Five of them each produce one document, and the next one starts only when I approve it. A sixth keeps the crew sharp:
  • Marvin, the scoper. Questions the idea until it’s clear what problem it solves, for whom, and what’s out of scope. Produces the brief.
  • Slartibartfast, the architect. Designs how it will be built: components, data, integrations, and the trade-offs behind each choice.
  • Eddie, the tester. Writes the QA plan: what has to be true for the build to count as done, written before anyone builds.
  • Trillian, the outcome owner. Defines what success looks like after launch, and stress-tests each measure so it can’t be gamed.
  • Ford Prefect, the planner. Breaks the approved design into small, ordered tasks, each tied to the tests that prove it.
  • Deep Thought, the calibrator. Reviews how the crew performed on real tickets and tunes them.
• Matters: decisions are made once, while they’re cheap to change.
• You’ve heard it called: spec-driven development, the opposite of “vibe coding”.

Swipe sideways to see the whole diagram

Tests locked up front

• Does: the tests that prove a feature are fixed before the build, and agents can’t weaken them.
• Matters: “it passed” means it passed the original bar.
• You’ve heard it called: guardrails.

Risk declared per product

• Does: every change is automatically classed, from routine to irreversible.
• Matters: risky work gets stricter handling by rule, not by mood.

2. Plan: “Where does the work live?”

One queue, graded at the door

• Does: every task lives in one tracker and is graded on arrival as runs-alone, needs-a-human or someday.
• Matters: nothing lives in anyone’s head.

Agents propose, triage decides

• Does: agents suggest follow-up work and a scheduled triage pass decides what’s real. Reviewer findings become tasks automatically.
• Matters: the backlog grows with real work, not agent noise.

3. Build: “Who does the work?”

A team’s worth of agents

• Does: a dozen or more agents work in parallel, with supervisor agents answering routine questions for me. There’s no fixed ceiling: the number is set by how fast changes can be reviewed, not by what the platform can run.
• Matters: one operator delivers a small team’s output.
• You’ve heard it called: multi-agent orchestration, agentic workflows.

The right model for the job

• Does: premium models on high-risk products, leaner ones on internal tooling, and more than one model vendor.
• Matters: spend follows value, with no lock-in to one vendor.
• You’ve heard it called: model routing.

Contained by design

• Does: agents run with scoped permissions, extensions are vetted before they load, and transcripts are scanned for leaked secrets.
• Matters: agents can’t do or leak what they shouldn’t.
• You’ve heard it called: least privilege.

4. Run: “Does it stop when I do?”

Works while I’m away

• Does: a dispatcher works the queue overnight and whenever I’m out, switched on and off from my phone. It runs several tasks at once, never two on the same file. How many is a setting matched to review capacity, not a platform limit.
• Matters: idle hours become shipped work.
• You’ve heard it called: background agents, autonomous agents.

Paces itself

• Does: it works in priority order, waits until blockers have actually landed, stays inside the plan’s limits, and restarts stalled work.
• Matters: costs stay capped and nobody babysits it.

5. Assure: “Can we trust what it built?”

No agent marks its own homework

• Does: code from one AI is reviewed by a different AI from a different vendor.
• Matters: one model’s blind spots don’t compound.
• You’ve heard it called: AI code review, LLM-as-a-judge.

Reviewers tested before they’re trusted

• Does: reviewers were calibrated on deliberately planted defects and caught 5 of 5. Three rival reviewers were trialled and rejected.
• Matters: reviewers are chosen on evidence, not marketing.
• You’ve heard it called: evals.

Every change carries its own evidence

• Does: each change states the problem, the outcome and the proof, and machines check those claims.
• Matters: approval is on outcomes, so nobody needs to read code.

6. Decide: “Where’s the human?”

Two touchpoints, by design

• Does: I approve the spec and the outcome, and irreversible changes always need my tap.
• Matters: human judgement goes where it counts: product, architecture and risk.
• You’ve heard it called: human-in-the-loop. It’s not a “lights-out factory”: two lights stay on, deliberately.

Decisions on my phone

• Does: questions arrive answerable in three lines, with tap-to-answer and one-tap approval to ship. A morning message lists only what needs me.
• Matters: work doesn’t wait for me to be at a desk.

Every decision logged and audited

• Does: every decision is recorded. I audit how often I take the agent’s recommendation; over one fortnight it was 77%.
• Matters: there’s a full audit trail, and proof of exactly where a human still changes the outcome.
• You’ve heard it called: AI governance.

7. Learn: “Does it get better?”

Learns overnight

• Does: lessons from the day become improved agent instructions overnight, reviewed like any other change.
• Matters: a mistake is fixed once, in the system.
• You’ve heard it called: agent memory, context engineering. It’s instructions, not model training.

Do it twice, automate it

• Does: anything done by hand twice gets automated.
• Matters: manual work trends towards zero.

8. Stay honest: “How do we know it’s working?”

Checks that can fail

• Does: every safety check is tested by deliberately breaking what it guards, and a weekly sweep finds anything built but never switched on.
• Matters: a green light means something.

Measured every week

• Does: output, rework, cost per change and decision load are measured against the baseline, including when they get worse.
• Matters: decisions about AI run on data, not enthusiasm.
• You’ve heard it called: observability, AgentOps.

How review works

Swipe sideways to see the whole diagram

1. Triage. Every change is classed by risk.
2. Quick first pass. A low-cost AI flags the obvious problems.
3. Independent review. Codex reviews every change; CodeRabbit also reviews the high-stakes products. Reviewers read; they never edit.
4. One pass, not a loop. Fixes are re-checked once. We measured this: when reviewers looped until “clean”, more than half the late serious bugs came from the fixes themselves.
5. Every finding resolved. Each one is fixed, waived by a human, or escalated.
6. Hard gates. Automated checks that fail closed decide whether a change can ship.
7. Ship. Routine changes are heading towards shipping automatically once every check passes. Irreversible ones always need a human.

The numbers

Six weeks, 11 September – 23 October 2026

The busiest week (5–12 October)

What stayed with me is the point. The 36 hands-on tasks were production go-lives, security settings on live systems, vendor accounts and billing, and physical machine work. That’s exactly the work an executive would want a human to own. Everything else, the agents did.

Measured honestly. The same measurement showed output rising faster than efficiency, and the biggest hold-up was a reviewer’s spending cap, not the agents. Each finding became a change to the factory. Measuring, finding the constraint and fixing the system is the product.

Swipe sideways to see the whole diagram

Next on the line

• Pacing against a weekly budget, so idle agents free their slot.
• Turning the concurrency dial up as more review moves to machines.
• A second build engine (Codex) for mechanical work, alongside Claude.
• Automatic model switching as plan allowances run down.
• Agents that wake when their work ships and finish the follow-up themselves.
• Crash recovery from an agent’s own transcript.
• Automatic shipping for routine changes once every check passes, shadow-tested first.
• Routine decisions answered by policy, proposed to me until it earns trust.
• Lessons harvested from the wider community and adopted by trial, not by install.

What this means if you lead technology

Eighteen years in the rooms where technology programmes succeed or stall. Now I’ve built the factory myself.