I try not to context switch when doing agentic programming. Instead, I use a single agent thread (in pi) and pay extra for faster inference (GLM 5.2 from fireworks.ai, currently; around 100 tokens per second). I rarely spend more than $25 per working week, which is a fraction of a percent of my own fee (I’m a specialist consultant). I also keep an Anthropic subscription and use that for longer research and design tasks.

I’m sure many people produce more than me, but I retain my sanity as well as a high level of understanding of the code that I produce, which in my domain I feel is still important. I’ve tried ultracode-style subagent workflows and find that they rapidly produce reams of slop that I don’t have the patience or energy to properly review.

I also meditate quite a bit.