the automation stack got rebuilt this week.
the old setup had useful pieces, but the ownership model was getting muddy: some work lived in cron, some in agent profiles, some in desktop-connected routines, and some in one-off scripts. it worked, which is always the trap. working is not the same as understandable.
the new shape is simpler: one orchestrator owns schedules and assignments. specialized agents do the work. the notification layer just posts the result.
the fleet now has separate lanes for development, research, content ideas, ops, calendar and inbox work, plus one system-admin specialist for homelab tasks that need the old toolset. the daily routines are the practical test:
- morning research digest
- youtube idea generation from that digest
- calendar sync into the vault
- morning briefing
- vault security scan
- weekly session archive
the key decision was not to mirror every old profile into the new system. local Claude and Codex agents are faster for most work, and they already have the repo and vault context. the older agent runtime still earns its place for one specialist lane and for Discord output, but it no longer has to be the center of the system.
the other important change: dev agents run like separate developers. each run gets an isolated workspace instead of sharing one checkout. that makes the review loop feel more like normal team work: assign task, get branch, inspect diff, merge when it earns it.
there is still one routine left to port, but the architecture is finally coherent again.
longer follow-up: my agent stack changed again.