A design practice run by one person still carries the jobs of a studio. Someone has to triage the work, sequence it, review it, merge it, sweep the pipeline, plan the week, and write down the decisions so they hold. When the team is one person, those jobs don't disappear. They land on evenings.
This did not start with design, and it was never about replacing it. I had no urge to hand over the part of the work I am for. It started with delivery: getting designed work into code without me as the bottleneck. Then the management around that delivery — run it properly through the tracker, with agents doing the triage, the review, the merging. The lightbulb came later: a system that could manage delivery could operate on my behalf — reporting, planning, and now, at the operator level, running aspects of the operation whole. Somewhere along that path the real question surfaced: could the practice be written down precisely enough that something else could run it? Writing it down turned out to be design work of the ordinary kind — naming, structure, boundaries — pointed at my own operation.
Starting small
The canon repo opened on 3 June 2026. On 4 June the first agent-opened pull request was merged — a small change to a skill file, built by one agent, reviewed by another, merged on approval. One ticket, end to end, no hands on keys. Everything since has been that loop, hardened and widened.
The unit of the system is the artefact: a markdown file that says what a behaviour is, when it triggers, and where it must stop. Skills describe repeatable behaviours. Tasks and routines put them on a clock. Agent profiles give collections of them a mission and a remit. Conventions record the rules for every tool the system touches. The agents read these files to work; this site is generated from the same files. There is one source of truth and it is prose.
What broke
The honest version of this story is a list of failures, each of which became a rule.
On 13 June, two build sessions claimed the same ticket at the same moment and opened competing pull requests. The fix was structural: ticket status became the concurrency lock. Todo means ready, In Progress means claimed, and no separate mechanism can drift out of step with the board.
Later that month a scheduled task ran stale instructions for days. The schedule's configuration carried its own copy of the prompt, the canon had moved on, and nothing could see the gap. The fix became the wrapper model: every scheduled job is a thin loader that reads its canon artefact at run time and follows it verbatim. Configs froze; the words became the system.
And early on, a builder and a reviewer bounced the same pull request between them politely and indefinitely. Now bounces are capped at two strikes before a human looks. The deeper rule underneath all three: an agent never gates its own work, and silence is the one failure mode that is not tolerated. Anything unresolvable stops loudly — Blocked, urgent, assigned to me, with a plain sentence saying why.
Naming the fleet
For the first weeks the agents were labels on tickets. Naming them — Relay, Forge, Prism, Scout, Pulse and the rest — was not decoration. A named role can hold a mission, a remit, boundaries, and declared failure modes. When a pull request merges, the label says which role did it and which role approved it. The names are deliberately abstract, and there are no human names among them: this is an addition to the ways of working, not a person replaced.
The six-layer model came last, not first. After enough artefacts existed, the structure was found by classification rather than invented up front. One test decides everything: does an artefact run, or is it read? Facts sit below the behaviour, rules sit above it, and the four layers in between do the work.
What changed
Mornings now start with a brief the system wrote. Overnight routines drain the backlog on cheap tokens. The pipeline sweeps itself from the inbox into the tracker. A weekly retro turns friction the agents observed into proposed improvements to their own instructions — proposals, never edits, because canon changes are mine to approve.
My job changed shape. Less doing every job; more gating the ones that matter. The practice got a property solo work rarely has: it accumulates. Every decision logged, every rule written, every failure converted into structure that outlasts the week it happened in.
What it is not
It is not autonomy theatre. A human owns validation, merges are gated, and the numbers published here are only the ones you can check against the public repo. Plenty is still unfinished: some artefacts are drafts and say so, some skills are procedure without craft, and the system's own measure of the attention it frees is not yet precise enough to claim.
Take it
All 94 artefacts are downloadable as plain markdown — the same files the agents run on. They were written for one practice, and the specifics won't fit yours. The ideas will: write the behaviour down, separate the doing from the gating, make every failure a rule, and keep one source of truth in words you can read.