From Agent Loops to Structured Graphs:A Scheduler-Theoretic Framework for LLM Agent Execution

Researchers propose SGH (Structured Graph Harness), a framework that replaces iterative Agent Loops with explicit directed acyclic graphs (DAGs) for LLM agent execution. The approach addresses structural weaknesses in current agent design by enforcing immutable execution plans, separating planning from recovery, and implementing strict escalation protocols, trading some flexibility for improved controllability and verifiability.

This arXiv position paper addresses a fundamental architectural problem in LLM agent systems that has grown more acute as enterprises deploy autonomous AI systems at scale. Traditional Agent Loops—where a single LLM iteratively decides next steps by processing an expanding context window—create implicit dependencies, unbounded recovery loops, and debugging nightmares when execution histories become mutable. The authors' scheduler-theoretic framework reinterprets this problem through classical computer science lenses, treating Agent Loops as single-unit ready schedulers where opaque LLM inference replaces transparent policy.

The proposal's practical significance lies in shifting control flow from implicit context into explicit, statically verifiable DAGs. By enforcing three core commitments—immutable execution plans within versions, layered separation of planning and recovery, and hierarchical escalation protocols—SGH sacrifices expressiveness for auditability and formal guarantees. This tradeoff matters particularly for regulated industries and safety-critical applications where black-box agent behavior becomes unacceptable.

For the AI infrastructure sector, this represents emerging consensus that production-grade agent systems require formal verification properties current implementations lack. The authors' trade-off analysis across 70 surveyed systems provides valuable taxonomy for developers choosing between flexibility and controllability. The experimental framework's seven-group design anticipates future empirical validation of these theoretical claims.

The framework doesn't solve non-determinism inherent in LLM nodes themselves, which remains the sector's fundamental challenge. Future work must determine whether structured graphs genuinely improve reliability metrics or merely move complexity from execution to plan generation. The paper's theoretical contribution is substantial even without production implementation.

• →SGH proposes replacing iterative Agent Loops with explicit DAGs to eliminate implicit dependencies and improve auditability in LLM agent systems
• →The framework enforces immutable execution plans and layered recovery separation, trading expressiveness for controllability and formal verification properties
• →Scheduler theory applied to LLM nodes reveals fundamental challenges around non-determinism that structured graphs alone cannot resolve
• →Trade-off analysis across 70 systems establishes that different agent architectures optimize for different combinations of controllability, expressiveness, and implementability
• →Position paper provides theoretical foundation and experimental protocol but lacks production implementation or empirical validation results