Loop engineering in the wild: 217 confirmed always-on agent loops, almost no committed state
Aug 22 Lulla et al.: gray-lit definition + mine of 36,710 repos; 217 confirmed autonomous loops (mostly event PR-review/cron triage); configs committed, runtime state almost never versioned.
Source: Loop Engineering: Building Blocks, Adoption, and Impact
Jai Lal Lulla, Vahram Nersesyan, Seyedmoein Mohsenimofidi, Christoph Treude, and Sebastian Baltes put loop engineering—designing systems that start coding agents on a schedule or repo event and stop them on a machine-checkable condition—on an empirical footing. A gray-literature review converges on building blocks: triggers, stop conditions, persistent state, independent verifiers, budgets, and escalation. A mining study then scans 36,710 engineered software repos and manually verifies every heuristic match.
They confirm autonomous agent loops in 217 of 256 candidate repos (0.59% of the frame; precision 0.868). Of those, 180 run on repo events only (mostly automatic PR review), 21 on schedule only, 15 on both. Claude Code dominates attributions (189/217). The split that matters for factories: committed triggers are common (253 repos with schedule/event agent workflows), but the discourse-prescribed state files are almost absent—two candidates, both rejected as homonyms; zero committed stop-condition phrase matches, budget files, or loop-referenced verifier subagents. Loops run; their runtime state lives outside version control (issue trackers, platform run histories). A planned autonomy experiment (interactive vs goal-driven vs scheduled) is outlined for the journal extension.
Why it matters for code review
Always-on factories are already shipping in open source—mostly as event-triggered review and cron triage—without the STATE.md orthodoxy the blogs prescribe. Review continuous agent lanes for what this study can actually see: committed schedules and event workflows that invoke an agent CLI, named escalation, and whether anyone can audit what the loop did when state never lands in git. FOMO for “overnight agents” without observable stop conditions and independent verification is exactly the unverifiable configuration the authors flag. Pair with Osmani’s loop engineering field note and LoopsBench’s dependency-DAG eval when promoting unattended merge-rate goals.