In June 2026, six words reorganized how software people talk about AI:
"design the loops that prompt your agents."
But "#loopengineering" didn't come from nowhere.
It's the fourth rung on a ladder we've been climbing for four years.
We climbed it because the AI kept changing shape underneath us.
#Prompt → #Context → #Harness → #Loop.
Each rung was forced by a change in what the model fundamentally was:
a #predictor, then a #reasoner, then an #agent, then an agent you can leave running.
The engineering is the shadow the technology casts.
- Scaling gave us a predictor → so we engineered prompts.
- Reasoning + tools gave us an agent → so we engineered harnesses.
- Autonomy gave us a self-running agent → so we engineered loops.
But look closely and every rung has been coping with the same absence:
the agent is brilliant, and then it forgets.
The loop got us to the very edge of it.
Its own trick is the reset: wipe the context, let the state live on disk. State outside the model. That's the clue.
The next rung isn't a bigger input or a faster loop.
It's to stop routing around the forgetting and engineer against it directly.
To treat the agent as something that persists and accumulates worth more next month than this month, because it's been running the whole time.
New essay series. Part 1
002548
