A spec-driven multi-agent engineering pipeline

At work I deliver every change the same way: spec, approved design, implementation, verification, updated spec. It works, but it lives in my head and my habits.

I wanted the same discipline to hold when AI agents do most of the typing, and to hold on a bigger product than one module.

One general-purpose AI assistant writing a whole product drifts. It makes decisions nobody approved, edits files outside its job, and produces tests that don't map back to what was asked for.

I split the work into 9 role-scoped agents: manager, change review, architecture, data, API, frontend, infra, security and QA. Each one owns a stage and writes only its own artefacts.

Between the stages sit 7 human approval gates. No stage moves on without a person approving it.

There are three tracks, so the process fits the change: a full feature, a light change, or a spike to explore an unknown first.

A write guard limits each agent to its own files, so the API agent can't quietly change the data model. An automated traceability check confirms that every acceptance criterion has a test before anything is released.

It's at v0.1 and in active use. It built a SaaS product's foundation, covering architecture, frontend, backend, tests and infrastructure, in about two weeks.

AI drafts, I decide, and the gates are where that line lives. Writing my own way of working down as rules a tool must follow also made me much clearer about why each step exists.