Built a 9-seat AI agent system to run my apps while I sleep. Here's what actually worked.

Running 6 apps and a content system while holding a full-time senior role.

The only way that works is if most of the system runs without me. So I built it to.

Nine specialized AI agents with defined scopes and veto conditions. Thirty-two memory files, six database tables, four edge functions. A verification agent who takes nothing on faith and has stop authority over the others. An asset engine that fans out in parallel instead of in sequence. More than 2,000 agent sessions per year, most overnight.

What worked:

Keeping memory in the database, not the model. Context windows forget. Databases don't. Twelve recurring failures codified as named patterns. I've solved each once.

Making verification a separate seat. The thing that builds will always declare itself done. Having a separate agent reproduce failures on real devices before anything ships has caught more regressions than I can count.

Building to adapter contracts. Every source plugs into one interface. When tools change (and they always do), nothing downstream breaks.

What didn't:

Aiming a competent agent at the wrong audience. She executed perfectly. Zero revenue. Bad problem framing upstream makes the best agent worse than nothing.

Trusting scaffolding. Generated iOS scaffolding has shipped wrong architecture, missing git commits, and permission strings that crash on real hardware. Pre-archive checklists exist now.

Letting optimization loops run unscored. Unbounded self-improvement is drift. Every loop has a threshold and a kill condition now.

Numbers: Six apps. Personal content system covering 40+ accounts. More than 2,000 agent sessions in the last twelve months.

Still figuring out: how to make the memory layer more portable across projects. Anyone else building multi-agent systems for solo products?