
I started by joining useful things together. The tools changed. The habit stayed. What grew around it was a way to direct the work, catch the misses and see what had actually survived.
The useful habit came before the agents.
The early stack was a reverse proxy, a VPS and WordPress plugins. I could join pieces that already did their jobs instead of rebuilding every capability. Composition was the skill: work out what each piece was good at, then make the pieces cooperate.
Plugins gave way to APIs, cloud services, infrastructure, models, skills and agents. The work moved from putting a website together towards operating the platform and its delivery. There was no clean line where a pair-programming tool became a factory. The coordination problem grew with the work.
Read the diagram in text
Compose tools: use plugins, APIs and services for the jobs they already do well.
Work in a pair, then delegate: one human and one agent become bounded tasks with explicit owners.
Parallel work brings competing changes. Independent review and qualification become necessary.
Factory operation ties intent and work to evidence about an exact release. This is a sequence, not a dated timeline.
Writing code stopped being the only bottleneck.
As agents worked across several products for extended stretches, the coordination problem changed. With that much happening at once, I needed to know more than whether a task had started. What had changed? Who owned it? Had another agent challenged it? Had it reached the product anyone could actually open?
A backlog only helps if it can answer those questions. Bounded tasks and reviewable revisions make parallel work understandable. Durable records let the next agent pick it up without asking a human to remember every conversation.

The point of the work is still the thing someone can use. In Bubbles, that means a game to play and a result to share. Play Bubbles ↗
The factory floor, 28 September 2026
The first reviewed portfolio put Bubbles in final convergence and Fighter in A+ convergence. Pirates was in source archaeology with Gate B locked. MAX needed proof on the actual walker. These were different finish lines, not four versions of “nearly done”. This is the historical snapshot; each project’s current record keeps its later review date.
Historical evidence, reviewed 28 September 2026. It describes maturity without asserting a green browser matrix or completed owner acceptance.
Inspect the original Bubbles recordThe historical record stays available beside the current work. Open the 28 September factory snapshot
The tests passed. The game still looked wrong.
Fighter exposed the gap between measuring an interface and judging it. Responsive geometry and screenshot checks could confirm containment and expected bounds while the composition still missed its purpose.
That failure changed the control. Visual hierarchy, whole journeys and copy needed their own review. The same distinction later informed Bubbles. Repeating a check that had already missed the problem would not have taught the factory anything.
Read the diagram in text
- Belief · State the modelRecord the expectation
- Attempt · Make the changeImplement against the model
- Reality · Observe the resultCheck the real behaviour
- Incident · Name the missIdentify what escaped
- Control · Change the systemPrevent repeat failures
The incident record keeps the belief, the miss, the cause and the change together. Read the Fighter incident

Sometimes the useful thing is a boundary.
Pirates is a different test. It is reconstructing Master of the Secret Sea, so a plausible invented rule is a mistake. The agents have to recover the source relationship, account for the evidence and compare the whole game before expansion earns its place.
A source gap stays open. It does not become permission to fill in something convincing. The shape of that decision matters as much as the speed of the implementation.
Read the diagram in text
Recover original source rules and preserve provenance.
Implement evidenced rules, then compare the integrated game against its source.
If a source rule is unknown, keep it open. Plausibility is not evidence.
Gate B remains locked in this example. Expansion requires integrated parity and the owning gate’s acceptance.
“Done” was hiding several different questions.
A build can show that code compiled. Review can challenge a change. Qualification can test a candidate. None of those observations tells me that the named version is live, that it meets the intended bar or that it works on the actual device.
MAX makes that distinction unavoidable. Software can advance while the real Pi, walker and FIND experience still need physical proof. A browser cannot supply it. The evidence has to say which claim it supports and where it stops.
Read the diagram in text
- Observed · A source or behaviour was inspectedEvidence has a place and a time
- Implemented · The requested change existsThe code or product state changed
- Merged · The reviewed change entered its branchThe integrated revision is identifiable
- Qualified · Required checks passed on the candidateThe named release gates were satisfied
- Production verified · The deployed version was observedThe live system matched the claim
- Separate question · Human acceptanceThe accountable human confirms it meets the intended bar; this is not granted by deployment.
- Separate question · Physical proofObserve the actual device, person and environment. Human taste acceptance cannot replace this evidence.
The original reviewed record explicitly leaves the physical product unproven.
Inspect MAX’s physical-proof boundaryNow the connections are something you can try.
Woodhouse makes the work easier to find, then lets curiosity go deeper. Games keep their own character. LoftwahFM supplies original music and a versioned streaming client. App Composer joins independently owned apps into a working home without making every part inseparable.
The useful question is what someone can do with it. Try the small game with FM music. Explore a guestbook connected to a creator’s collection. Inspect a source, use a starter and keep ownership of the result. Working examples are available; independent adoption and demand still need their own observations.
The factory is the continuation of this story. Try the working connections in Factory →
Or start with something to play, listen to or build. Explore the projects →
The public figures show recorded activity and the limits of its collection. Inspect Analytics →
Source and review
Founding account and historical portfolio reviewed 28 September 2026; expanded editorial presentation reviewed 18 October 2026. The original account supplies a sequence, not dates or verbatim conversations. Product captures retain their own dates. Current project records remain linked separately from this history. The publication does not reproduce private repository bodies or conversation transcripts.