Agent Harness Engineering vs Loop Engineering vs Graph Engineering
A practical guide to the three architecture layers people keep mixing together
You’ve been told your agent is unreliable because you picked the wrong model.
You didn’t. You picked the wrong layer.
I know because I did it first. Upgraded the model. Same failure. Now more expensive.
Last week a founder ripped out Claude, swapped in a “smarter” model, and got the same broken agent back. The failure was never in the model. It was in the machinery around it - three layers everyone keeps blending into one buzzword soup.
Agent harness. Loop engineering. Graph engineering.
I spent the week pulling apart a research agent that kept lying about being finished. By the end of this issue, you will get the exact diagnostic I now use to fix any agent in under a minute - plus a free Agent Failure Diagnostic Cheatsheet.
By the end of this, you’ll stop paying for a smarter model to fix a problem the model was never causing. You’ll look at a broken agent - one that lies about being done, burns tokens in circles, or falls apart across sessions - and know in thirty seconds which of three things to fix. Not guess. Know.
That’s the difference between someone who uses AI and someone who engineers it.
