This page describes how the agent-aaron.ai programme is organised and how it runs. One private repository holds the whole programme, with each project a folder within it, and the person and the agent that work on it are each defined by what they may do there. Below that are the flows the work moves through: the loop every change goes round, the way the programme corrects itself as part of that loop, and the two flows built on it. The operating model is itself part of the work being shown, because the programme runs on the patterns it studies.
Everything the programme holds lives in one private repository. Master is its top level, and each project is a folder within it with its own CLAUDE.md: the instructions that say what the project is for and how a session works there. A person directs the work and an agent does it. Select anything to see what it reads, what it writes and what gates it.
The repository above is structure. The work moves through it in a small set of repeatable flows. The loop is the one every change goes round, and the others are built on it.
Every change the programme makes goes round this loop. The colours show who acts at each step. Select a step to read what happens there.
The programme is built to catch and correct its own failures, including the assistant's, and to carry forward its own ideas. Neither is a separate process that someone has to remember to run: every step is part of the session loop, so it happens in every session.
Failures are recorded as they happen, with nothing filtered out, and are sorted afterwards. A repeat is marked as recurring rather than recorded as something new, so a pattern shows up as a pattern, and a minor one-off that has not come back within thirty days is closed.
A failure counts as fixed only when the fix reaches the place where the failure happens and something observable would show it happening again. Writing a rule down is not the same as fixing it, which is why the programme prefers checks in code to instructions.
Ideas follow a shorter loop. The weekly reports and the source and tool assessments each propose a small handful of things the programme could adopt or change; nothing writes one down on its own judgement. Aaron reads the proposals and any unclaimed backlog at the start of a session and decides what is worth carrying forward, the same way he decides which failures are worth fixing.
The report skill can also write a retrospective on the operating model as a whole.
How anything reaches the public. It is the loop with steps added before and after it, shown dashed.
A change to copy alone skips the structure step. The project a piece comes from owns what it says. Publishing owns how it is presented and the act of publishing it, and writes the copy in the programme's public voice.
The programme's weekly information engine, built on a simple principle: dumb retrieval, smart selection.
What the pipeline writes goes straight into the repository. Changes to the pipeline itself go through the loop like any other code.