About this workbench

I've been building products for a long time. Recently I turned myself into a one-person-plus-agents product studio: I make the product calls during the day, a fleet of agents turns those calls into code at night, and a second fleet verifies the first one didn't lie to me.

This blog is the studio's public workbench. Four series, four production lines:

• Into the Harness — the design, scars, and methodology of agent harnesses: context engineering, tool permissions, acceptance gates, memory systems, eval loops.
• Studio Logs — shipping products as one person with a fleet of agents: real projects, real numbers, real crashes.
• Study Notes — close readings of papers and engineering write-ups, reorganized into an evidence system.
• AI Watch — observations and experiments on model behavior, the AI ecosystem, and how humans and agents work together.

The writing principle is the same as my agent principle: never trust self-reports; demand evidence. Every post aims to be spec'd (what question it answers), evidenced (real code and data), and bounded (what is inference, stated as such).

Every post ships in two native languages — written separately, not machine-translated. English lives here; the Chinese originals are at the root. Posts are syndicated to X, DEV.to, Medium, and Chinese platforms by an agent skill — the pipeline itself is part of the story.

Find me: GitHub @Octo-o-o-o · RSS