Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-25 · high
A gray-literature review defines the pattern as triggered runs with a machine-checkable stop condition, then tests it against 36,710 repositories.
Why now: Practitioners started using the term in June 2026, and this review is what turns it into a count of real repos.
A paper on arXiv calls the layer of agent design above prompt and context engineering "loop engineering."
That distinction matters because most things called agent loops get no check for whether the work behind them is real. The review mined 36,710 repositories, flagged 256 as candidates, and confirmed only 217 running an actual operating loop.
The recipe it converges on starts with triggered runs, kicked off by a schedule or a repo event, that stop only when a machine-checkable condition holds. Each loop keeps a persistent state file the agent reads and writes between runs, and a verifier sub-agent checks the work before the loop closes. A token budget and a defined human escalation point close out the design.
Practitioners started using the term in June 2026. Whether 217 confirmed loops is a lot depends on what share of the 36,710 mined repositories were even attempting agent automation to begin with, and the paper doesn't say.
Each link below shares sources, entities, or timing with this story.
Loop uses Claude / Same source domain / Shared topic / Tension
Linked by a graph relationship (Loop uses Claude); reported by the same outlet (arxiv.org); overlapping topics (agent, document).
Loop uses Claude / Same source domain / Shared topic
Linked by a graph relationship (Loop uses Claude); reported by the same outlet (arxiv.org); overlapping topics (agent, context).
Loop uses Codex / Same source domain / Shared topic
Linked by a graph relationship (Loop uses Codex); reported by the same outlet (arxiv.org); overlapping topics (agent, context).
Linked by a graph relationship (Loop uses Codex); reported by the same outlet (arxiv.org); overlapping topics (agent, context).
Loop uses Claude / Same source domain / Shared topic
Linked by a graph relationship (Loop uses Claude); reported by the same outlet (arxiv.org); overlapping topics (agent, context).
Loop uses Codex / Shared entity: Loop / Shared topic / Earlier coverage
Linked by a graph relationship (Loop uses Codex); both cover Loop; overlapping topics (agent, engineering, loop).
Loop uses Claude / Shared entity: Loop / Shared topic / Earlier coverage
Linked by a graph relationship (Loop uses Claude); both cover Loop; overlapping topics (agent, engineering, loop).
Loop uses Claude / Shared topic / Tension
Linked by a graph relationship (Loop uses Claude); overlapping topics (agent, engineering); pushes against this story (but).