AOI: Context-Aware Multi-Agent Operations via Dynamic Scheduling and Hierarchical Memory Compression

1. Problem Statement

Modern IT infrastructures have evolved into highly distributed, cloud-native systems characterized by microservices and dynamic orchestration. While this architecture enhances scalability, it introduces unprecedented complexity and volatility, leading to four critical challenges in automated operations (AIOps):

• Information Overload: The exponential growth of logs, metrics, and alerts overwhelms traditional analysis pipelines, degrading decision accuracy and increasing latency.
• Poor Task Coordination: Existing systems lack intelligent task decomposition and dynamic scheduling, leading to inefficient resource utilization.
• Context Loss: Traditional approaches fail to preserve crucial context during long-running operations, impairing diagnosis continuity and repair consistency.
• Limited Adaptability: Most models cannot evolve with system drift or new failure modes, limiting their ability to generalize to unseen conditions.
• Safety Risks: Automated remediation often lacks robust safety mechanisms, risking destructive actions during fault recovery.

2. Methodology: The AOI Framework

The authors propose AOI (AI-Oriented Operations), a novel multi-agent collaborative framework designed to enable scalable, adaptive, and context-aware autonomous operations. The framework integrates three specialized agents, an LLM-based context compressor, and a hierarchical memory system.

A. Multi-Agent Collaborative Architecture

AOI employs a three-agent structure to ensure clear responsibility separation and operational safety:

1. Observer Agent: Acts as the central coordinator. It performs hierarchical task decomposition, analyzes system states, and implements Dynamic Task Scheduling. It adaptively prioritizes between "probing" (information gathering) and "execution" (remediation) based on real-time system states and historical patterns to optimize the Mean Time to Resolution (MTTR).
2. Probe Agent: A strictly read-only agent responsible for safe information gathering. It adheres to the principle of least privilege, executing only non-destructive queries (e.g., log reads, database introspection) to validate hypotheses without altering system states.
3. Executor Agent: Responsible for system modification and remediation. It operates under strict safety constraints, including checkpointing before execution and automatic rollback mechanisms if failures are detected, ensuring a "fail-fast, recover-quickly" paradigm.

B. LLM-Based Context Compression

To address information overload, AOI introduces a Context Compressor module:

• Mechanism: It utilizes a sliding-window mechanism with a 50% overlap ratio to process raw operational data.
• Function: Leveraging Large Language Models (LLMs), it performs semantic-aware summarization. Instead of simple keyword filtering, it identifies and preserves critical operational information (error patterns, causal relationships, thresholds) while discarding redundant data.
• Theoretical Guarantee: The paper provides a theoretical proof (Theorem 2 & 3) showing that this compression is practically stable and preserves task-relevant mutual information up to a bounded error, preventing unbounded semantic drift.

C. Three-Layer Hierarchical Memory

The framework manages data lifecycle through three distinct layers:

1. Raw Context Storage: Stores unprocessed data from Probe/Executor agents (24-hour retention).
2. Task Queue Management: Maintains structured execution instructions and dependencies.
3. Compressed Context Cache: Houses LLM-processed strategic information for long-term decision-making (7-day retention), enabling efficient retrieval and cross-task reuse.

D. Theoretical Foundations

The authors formalize the coordination problem as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP). They provide theoretical proofs for:

• Dynamic Scheduling Convergence: Showing that the adaptive scheduling strategy converges to the optimal static policy with a bounded regret.
• Context Compression Stability: Proving that the compression operator is contractive, ensuring the compressed context remains bounded around an equilibrium state.
• Information Preservation: Establishing bounds on the loss of task-relevant information relative to window size and overlap ratio.

3. Key Contributions

1. Multi-Agent Architecture: A specialized triad of Observer, Probe, and Executor agents that separates coordination, safe observation, and risky execution, significantly enhancing system safety.
2. Intelligent Context Compression: An LLM-driven sliding-window compressor that achieves high compression ratios (>70%) while retaining >90% of critical diagnostic information.
3. Dynamic Task Scheduling: An adaptive algorithm that balances exploration (probing) and exploitation (execution) based on real-time uncertainty and system states.
4. Hierarchical Memory System: A three-layer architecture designed for efficient storage, retrieval, and reuse of operational context across incidents.
5. Theoretical Guarantees: Formal analysis of convergence, stability, and information preservation bounds for the proposed framework.

4. Experimental Results

The framework was evaluated on synthetic benchmarks (AIOpsLab) and real-world datasets (HDFS, BGL, OpenStack) against four baselines: Rule-based Expert Systems, Traditional AIOps Pipelines, Single-Agent LLMs, and Baseline Multi-Agent Systems.

Key Performance Metrics:

• Task Success Rate (TSR): AOI achieved 94.2%, outperforming the best baseline (B-MAS) by 7.8% and rule-based systems by 26.4%.
• Mean Time to Resolution (MTTR): AOI reduced MTTR to 22.1 minutes, a 34.4% improvement over the best baseline (33.7 minutes).
• Context Compression: Achieved a 72.4% compression ratio while preserving 92.8% of critical information (Information Preservation Score).
• Safety: AOI demonstrated the lowest False Positive Rate (3.1%) and the highest System Safety Score (96.7%).
• Scalability: The system maintained stable performance (TSR > 88%) and near-linear resource utilization efficiency even with up to 50 concurrent tasks.

Ablation Studies:
Removing the Context Compressor dropped TSR by 5.7% and increased MTTR by 7.7 minutes. Removing Dynamic Scheduling increased MTTR by 20.8%, confirming the necessity of adaptive prioritization.

5. Significance and Impact

• Paradigm Shift: AOI moves beyond static rule-based or isolated ML approaches toward a context-aware, collaborative, and adaptive autonomous operations model.
• Safety-First Automation: By strictly separating read-only observation from write-capable execution and enforcing checkpointing, AOI addresses the critical safety concerns that have hindered the adoption of fully autonomous IT operations.
• Scalability: The framework demonstrates the ability to manage the complexity of next-generation cloud-native infrastructures where human intervention is no longer feasible due to data volume and system volatility.
• Generalization: The model shows strong transferability across diverse datasets (storage, microservices, cloud platforms), indicating it is not overfitted to specific log formats but understands underlying operational semantics.

In conclusion, AOI presents a robust solution for the "alert storm" and complexity of modern IT, offering a path toward truly intelligent, self-healing, and safe infrastructure management with minimal human intervention.