339 agent sessions last week, one boxMany agents, self-healing infra, metrics over code.
Claude Code sessions on one box (life), 27 Oct to 3 Nov, from its own telemetry. The fleet has more boxes than this one.
Kiter when the wind is right.The factory runs itself in the meantime, with a little oversight.
The last 24 hours on that box, 2 to 3 Nov, 09:00 to 09:00 Oslo time. 10 of its 48 scheduled jobs, one tick per hour of runs.
Grey ran fine. Red failed. Yellow failed and came back (2 in this window, 0 still red at the hour). 14,298 job runs in all. Source: the box's job telemetry, read 3 Nov.
The posts live on LinkedIn, not here. Read them there. I post about once a week.
Running many agents and services without care and feeding.The basement server, ClickHouse, the fleet controller.
What good engineering means when change is cheap.Best practice, tech debt, over-engineering, when not to optimise.
Concrete things I built, what broke, and what the data said.
The consultant habit of chasing tech instead of shipping product.I had it for years.
Nothing posted here yet.
I used to be scared of building. Not the building itself, the being wrong afterwards. Change was expensive, so I spent more time controlling chaos around my code than shipping the product.
That block is gone. I run an agent factory from a server in my basement: a fleet of AI agents and services with their own metrics, monitoring and self-healing, so I spend my time on what gets built rather than on keeping it alive. The most useful tool I own is a fleet controller that is one HTML file past 10,000 lines. It is not best practice. It works.
Before this: CTO at Diwala, consultant at Bekk, CS at NTNU.