Overview: the map before the deep dives

Part 1 of Jarvis: the architecture series by Jack Stovell · published 2026-07-29

Overview: the map before the deep dives — architecture diagram

I accidentally built 22 agents for a to-do list. Everything after that was deliberate.

It's called Jarvis. Bit much, I know.

The diagram attached is the honest version of what's running. Not a monolith, peer to peer agents, each one its own process, its own repo, its own restart policy. If the news agent falls over at 3am, voice doesn't even notice.

The bus is the bit I'm proudest of. It's how the agents talk to each other — machine to machine, no human in the loop. Discord channels as the primary transport, because it's free, observable, and every frame is replayable. The audit trail IS the transport. Postgres LISTEN/NOTIFY sits underneath as automatic failover, because relying on someone else's uptime for your own nervous system felt like a bad bet. Every frame's HMAC-signed.

One rule for this whole series: every number comes from the live system, not a doc. systemctl, psql, the actual latency logs. If a figure shows up, it carries the date I pulled it.

One box also means real constraints. 4 vCPUs doesn't forgive you. Cheap-model tiering, hard cost caps, splitting work properly, the same discipline that matters at scale, just smaller.

Architecture write-up is public on GitHub. More coming on the bus, voice latency, the memory graph, and what broke along the way.