Prompt: "reason step by step"
text-to-image
Turning research into products people actually use.
Picked things I'm genuinely into.
Thoughts on AI, leadership, and where this is all going.
The tools I reach for every day.
Git versions files, but a coding agent accumulates processes, packages, sandbox state and a warm cache that a checkout can't restore. Shepherd makes every agent step a reversible commit, so a supervisor can catch bad changes before they land.
Cutting tool-output tokens can raise your AI agent bill instead of lowering it. Why prompt caching means token reduction isn't cost reduction, and what to measure instead.
Harness engineering is where AI agent capability actually lives. Context management, tools, memory and a real workspace decide how much of a model's intelligence you get, explained through four interactive demos.
