How Building 7 Apps Changed How I Think About Architecture

inbox0 is a 3-agent pipeline: Fetcher summarizes, Analyst categorizes, Writer drafts replies. I was not designing a CRUD service with a thin AI wrapper. I was designing how agents hand off work, what they produce, and how they fail. Different mental model.

Every app works standalone. No shared database. No service mesh. But with six apps running and none talking to each other, I had no single place to ask about them. So I built Gopher: a seventh app that owns no data of its own. It queries the others and answers in plain language.

The steward of the estate, not the owner of the rooms. Sovereignty was the design principle. A consent router was the emergent requirement.

Scribe (live transcription) sounded simple. What followed: Safari optimizing away the audio callback, ASGI dicts handling WebSocket binary differently across browsers, Whisper blocking the event loop, a GC bug dropping the final transcript on stop.

That gap is not a debugging problem. It is an architecture problem debugging reveals.

DrawmaticAI had OAuth, CSRF tokens, invite gates. A proper review returned 13 findings. The worst: IDOR on every diagram endpoint. Any authenticated user could read, modify, or delete another user's diagram by changing a UUID. One ownership check across 16 endpoints fixed it.

Obvious in retrospect. Embarrassing in practice.

One Ollama instance handles email triage, transcription, diagram generation, code review, and meeting summaries across all seven apps. The economics crossed the threshold. Nothing leaves the hardware.

You are not designing how data flows through a system. You are designing how agency flows: which agents decide what, on whose behalf, with what rollback and audit trail when they get it wrong.