Addy’s piece on Loop Engineering is excellent.
But I think the term is easy to flatten into:
“let the agent run automatically.”
That misses the point.
Loop Engineering is not automation.
It is system design.
The old workflow was:
prompt → output → prompt → output
The human stayed inside the loop, steering every step.
The new leverage is different:
design the loop once,
then let the system handle the repetitive steering.
A useful agent loop needs four things:
1. feedback
2. verifier
3. memory
4. state
Without them, you don’t have a loop.
You have repeated generation.
Verifier is the part most people skip.
A model grading its own work is structurally weak.
Better pattern:
generator does the work
verifier checks the artifact
human owns the final judgment
“Done” should not mean “the agent said so.”
Memory is not chat history.
Useful memory looks more like:
fail → investigate → verify → distill → consult
If a failure does not become a rule the next run can use, it did not compound.
It just became another note.
Afterall, the goal of Loop Engineer is not to remove human judgment.
It is to stop wasting human judgment on repetitive steering.
Loop Engineering 真正的重点是:
1. 环境能不能给反馈
2. 有没有独立 verifier
3. 失败能不能沉淀成下次会用的 memory
没有这三件事,loop 只会是自动化地制造 "slop"