-
How I run 19 agent sessions at once
A coding-agent session is worth treating as a long-lived named worker rather than a process you start per task. Nineteen of mine run permanently on one headless machine, reset in place with /clear and re-pointed with a slash command. Here's the addressing scheme, what forces a real restart, and the two places it falls over.
agentstoolingprocess
-
Vibe coding is not binary
The vibe-versus-serious split measured how much code you type, and that has stopped being a useful signal. Across 37 active days and 77,720 agent actions, 12% authored code and 68% ran commands and read output. The line that still means something is not typing volume - it is who owns verification and who owns the consequences.
agentsprocesstooling
-
Handing the easy tickets to a free local model cost 1.65x more
I built a harness to grind the mechanical backlog tail on a local Ollama model and measured what it actually saved. Across three real tickets the local-model path cost 1.65x the Claude tokens of just doing the work, and every output needed fixing before it could merge.
agentstoolingprocessgo
-
Dogfooding is a brilliant bug-finder and a terrible product manager
Every tool here is used to run the place, including the one reviewing this draft. That catches a class of defect nothing else does. It also produces a convincing-looking roadmap built entirely from the preferences of the least representative user available.
productprocesskythene