Skip to content

Foundations

The architecture I would recommend for a real engineering organisation looks roughly like this.

Kroki

Read top to bottom, the shape of the system is: intent becomes a dependency graph, the graph is allocated to isolated execution contexts, those contexts produce reviewable change sets, and a single controlled gate lets change into trunk.

The fan-out is deliberate and so is the fan-in. Everything above the merge queue is parallel; everything at and below it is serial.

The two halves of the problem

Half Mechanism Where it's covered
Parallel without collision Isolated environments, ownership domains, architectural decomposition Isolation, Ownership, Decomposition
Serial without becoming a bottleneck Small PRs, automated evidence, merge queue Integration

Most tooling in 2026 is strong on the first half and thin on the second — and thinnest of all on the part between them: deciding who is allowed to write what, right now. That gap is the subject of the Ownership section.

Nothing here is model-specific

Every mechanism on this site works the same whether the writer is Codex, Claude Code, OpenHands, or a human being with an IDE. That is the point — see Five Primitives.