Four phases of LLM control planes
We present a four-phase taxonomy for control planes that interface with large language models (LLMs) in software engineering. The framework begins with prompt engineering, which focuses on manually steering model behavior through instructions. It then progresses to context engineering, where relevant information is curated before entering the model's context window. The third phase, harness engineering, introduces multi-agent orchestration using rigid, hand-authored routing graphs. The final phase, loop engineering, is defined as a declarative goal combined with a recursive, self-managing control loop that continues until an independently verified completion condition is met.
Formalizing loop engineering
We formalize loop engineering as a discrete-time control system and identify six structural mechanisms that distinguish it from a conventional harness. These include a maker-checker split based on deterministic environmental signals, an externalized, disk-persisted state, and validation gates driven by compiler and test outcomes rather than by self-assessment. Additional mechanisms include structural delta-rollback, iteration compression, and a meta-loop that can modify the primary loop’s configuration. Together, these elements enable greater autonomy, reliability, and adaptability in LLM-driven systems.
Multi-agent detection and research scope
To improve system robustness, we formalize two loop-detection mechanisms: static cycle detection using strongly connected components (SCCs) and a proposed dynamic detector for implicit deadlocks based on state-embedding similarity and stagnant error patterns. This paper synthesizes and formalizes these mechanisms rather than claiming priority for either. Independent research has converged on several of the same concepts, including SCC-based loop detection, which was recently validated across 6,549 repositories. For each mechanism, we distinguish among approaches already supported by existing systems, areas where our formulation differs, and proposals that remain unvalidated, providing a structured framework for understanding the evolution from prompt-based interaction to autonomous, self-managing loop architectures.
