Reinforcement learning (RL) has shown promise for automating cloud resource scheduling, yet conventional RL-based schedulers often lack interpretability, violate service-level objectives (SLOs), and introduce substantial deployment risks in Kubernetes environments. In this paper, we propose a hybrid framework that integrates large language models (LLMs), safe reinforcement learning, and a digital twin simulator for distributed cloud resource scheduling. Rather than using the LLM as a direct decision-maker, our approach leverages it to generate scheduling constraints from SLO documents and historical cluster events, adjust reward functions, and produce human-readable explanations. A Proximal Policy Optimization agent with Lagrangian relaxation (PPO-Lagrangian) executes scheduling actions—including pod placement, autoscaling, and resource quota allocation—subject to the LLM-derived constraints. Before deployment, candidate policies are pre-evaluated in a digital twin simulator synchronized with the target cluster, and only policies passing safety verification are applied. We evaluate the framework on Google Online Boutique deployed on a Kubernetes cluster simulator under periodic, burst, and fault-injection workloads. Compared with five baselines, our approach reduces SLO violation rates by 39–68% and lowers normalized resource costs by 14–34% while maintaining competitive response times. Ablation studies confirm the contribution of each component to overall performance.

Large Language Models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks through multi-agent collaboration. However, existing approaches primarily focus on open-ended reasoning or simple task decomposition and lack structured frameworks for long-horizon decisionmaking with explicit planning, execution, and supervision. To address this, we propose Chain-of-Agents Planner (CoA-Planner), a framework that explicitly coordinates Planning Agents, Worker Agents, and a Supervisor Agent through a structured task graph. The Planning Agent decomposes complex tasks into a directed acyclic graph (DAG) of subtasks with dependencies,Worker Agents execute specialized subtasks, and the Supervisor Agent monitors progress and resolves conflicts at critical milestones. Our framework introduces adaptive role reconfiguration that dynamically adjusts agent assignments based on historical performance metrics. We evaluate CoA-Planner on several long-context benchmarks including document analysis (50K-120K tokens), codebase review, and technical report generation. Experimental results show that CoA-Planner achieves an average success rate of 67.9%, outperforming single-agent LLMs by 23.5 percentage points (pp), Chain-of-Thought with RAG by 17.9pp, and existing Chain-of-Agents frameworks by 10.4pp. The framework demonstrates efficient scaling with 4-5 worker agents and reduces error rates across all categories, particularly in context coverage and reasoning consistency.