Back in March this year, we rolled out the first iteration of our full-toolchain Operations Agent 1.0. At that time, we carried out extensive trials and explorations centered on the Five-Tier Operations Agent Architecture: Perception, Planning, Memory, Reasoning, and Execution. We also encountered numerous pitfalls along the way. For instance, we once held overly high expectations for large language models, assuming that sufficiently powerful models could resolve most issues. Previously, we underestimated the complexities of interconnection, constraint management and control across operations toolchains. This led us to a clear realization: enterprise-grade operations agents are far more than simple dialogue tools. They must operate within a controlled, trustworthy, auditable and sustainably iterative management framework, which we define as the Operations Agent Harness Paradigm. This paradigm empowers agents to take action while confining them within defined boundaries; it enables them to grasp on-site conditions without spiraling out of control.

From early spring to midsummer, we have continuously refined and perfected the Lewei Operations Agent Harness Paradigm. Meanwhile, after hundreds of days of polishing, we have delivered three core deliverables: QwenPaw for private deployment scenarios, ArkClaw integration for cloud environments, and Agentic Ops built upon Lewei CoT Chain-of-Thought.

Part 1 Two Architectures Address Enterprises’ Two Core Demands: Security & Control, Lightweight Cost Reduction

Many enterprises face a persistent dilemma when selecting AI operations solutions: Groups handling confidential data and subject to strict data compliance regulations cannot expose internal metrics, assets or work orders to the cloud; small and medium-sized businesses running cloud workloads cannot afford the high hardware and maintenance costs of private deployments. Instead of adopting a one-size-fits-all approach, we have developed two fully-fledged architectures deeply integrated with the Lewei Operations Agent to cater to both core requirements.

01 Private Deployment Solution – QwenPaw: Local Closed-Loop Eliminates LLM Operations Hallucinations

• Complete underlying operations closed-loop: For government, financial, manufacturing and other scenarios with strict internal control requirements, data must never leave the internal network. Built on the native five-tier architecture of the Lewei Operations Agent (Perception, Memory, Planning, Execution, Brain), QwenPaw embeds Perseus full-domain collection, CMDB asset management, monitoring/log/APM modules, and the Lerwee Claw execution engine. It natively delivers professional operations domain knowledge graphs, fault root cause analysis (RCA), alarm convergence, business topology analysis, and change control capabilities.
• Trusted data foundation: It does not rely on external public domain data for reasoning. All raw IT metrics, full logs, business topologies, historical fault work orders and operations scripts are stored locally. Every root cause analysis and self-healing solution is traceable, fundamentally resolving the hallucination issue where general large language models fabricate arbitrary fault conclusions.
• Closed-loop execution permissions: The agent can directly trigger restarts, capacity expansion, configuration modifications, inspections, work order dispatch and self-healing operations, with full access to the permission domain of enterprise IT systems.
• Unified multi-terminal entry: Previously, upon receiving alerts for high disk usage or CPU saturation, operations staff needed to log into dashboards, access hosts, manually run commands and clear caches line by line. Now, users can send alerts within DingTalk, Feishu or WeCom chat windows to issue instructions, receive alarms and view reports directly in the conversation interface.
• Multi-Agent task orchestration: Complex operations demands can be broken down to form collaborative agent teams dedicated to inspections, report generation, change audit and fault review documentation. The Lewei agent schedules these agents to complete batch and periodic composite tasks.
• Long-term memory + proactive push: The system retains operators’ interaction habits, common business scopes and duty rosters. It automatically retrieves Lewei data at scheduled intervals to generate daily/weekly reports and fault reviews, then pushes them to relevant personnel. Memory is synchronized across all channels, maintaining consistent context across all conversations.
• Supplementary general office capabilities: Native support for Office document parsing, report generation, PDF review creation and data summarization, converting raw operations outputs into standardized deliverable materials for external circulation.

Scenario Example: Alarm Diagnosis

• Alarms are triggered when disk space utilization exceeds 90%, with AI analysis and diagnosis

• Provide root cause analysis and self-healing solutions: alarms can be remotely handled with a single confirmation

IT Ops Agent Plugin available for download

Compatible with Lewei Operations Agent v8.2.1 and above

02 Cloud Solution – LerweeClaw: Deep Integration with VolcEngine, Out-of-the-Box, Compliant & Secure

Earlier this year, we trialed OpenClaw but encountered limitations including insufficient stability, inadequate compliance capabilities and poor adaptability to operations scenarios. We therefore established in-depth cooperation with VolcEngine’s ArkClaw to build a dedicated linkage architecture for LerweeClaw.

• Fully managed Claw instance services available upon deployment, capable of accommodating millions of endpoint connections in one stop. The product natively integrates four core capabilities: full-scenario compatibility, real-time monitoring and alerting, intelligent fault self-healing, and seamless version iteration. It eliminates repetitive operations overhead end-to-end and delivers lightweight enterprise operations.
• A mature architecture and automated operations system form a robust foundation to guarantee stable, secure and reliable long-term business operation. Leveraging near-instant instance startup and shutdown, the platform enables elastic scheduling of massive computing resources and self-service time-sharing reuse for users. It maximizes resource utilization efficiency through optimized resource scheduling mechanisms. Paired with refined end-to-end Token governance and dedicated performance optimization systems, dual optimizations via algorithms and scheduling directly cut Token consumption costs by half.
• A fully intelligent end-to-end operation loop covers environmental perception, solution planning and operational execution. It flexibly coordinates resources and computing power across diverse endpoints and heterogeneous systems, autonomously receiving and fulfilling user requests via self-operating mechanisms. The kernel dynamically upgrades alongside advances in artificial intelligence technology, helping the product build differentiated industry-leading advantages.
• The Lewei Operations Agent adopts chip-level encryption to deliver end-to-end full-link data encryption covering data collection, transmission and cloud storage. A trusted computing environment safeguards LLM operation security on the cloud. Real-time monitoring of model inference data governs compliance for all AIGC-generated content, with a dedicated Trust Center conducting full-lifecycle verification to achieve full transparency and trustworthiness across data, models and generated content.

Scenario Example: Fault Diagnosis, Analysis & Remediation

Part 2 Agentic Ops: Put an End to “Copy-Paste AI Interactions” for Operations Staff

In the past, AI-powered operations analysis involved cumbersome workflows: logging into the operations platform to retrieve metrics, exporting logs, copying asset information and pasting data into dialogue windows to submit queries. Each extra step introduced risks of data omission or error.

• Agentic Ops, built on Lewei CoT Chain-of-Thought, completely redefines AI interaction workflows for operations.
• Dedicated data placeholders are embedded in the editor to automatically pull CMDB assets, monitoring metrics, alarms and full-domain collection data to populate prompt templates instantly, eliminating manual data export and organization entirely.
• Over 90% of daily high-frequency operations scenarios – cluster load analysis, alarm storm consolidation, periodic resource inspections, business health weekly reports and fault root cause reviews – can be solidified as permanent standardized prompt templates. There is no need to rewrite instructions repeatedly; users only select target business clusters, hosts or services to trigger coordinated QwenPaw orchestration and Lewei operations decision-making, enabling one-click intelligent analysis. This marks a fundamental shift from rebuilding dialogue workflows from scratch every time to reusing accumulated operational knowledge.
• A split-screen visual editing mode enables real-time authoring and previewing. The left pane manages prompt template logic and data rule configuration, while the right pane synchronously displays the volume and details of connected real-world operations data, alongside the complete instructions submitted to the large language model. Template validity can be verified without test submissions, preempting parameter errors, missing data and ambiguous instructions to drastically boost template creation efficiency and accuracy.
• Pre-built mature operations scenarios are embedded within the platform: low-frequency business assessment, idle server identification, storage disk health inspection, and self-inspection of the operations agent itself, among others. One-click execution delivers comprehensive analysis and standardized, actionable governance plans. The platform even features agent self-diagnosis capabilities, automatically evaluating monitoring coverage scale, alarm pressure, server storage bottlenecks and persistent middleware disconnection issues to proactively warn of superimposed triple-risk hazards. This empowers operations teams to manage not only business faults but also the operations platform itself.

• Continuously Updated Scenario List

• Custom Chain-of-Thought Execution

• Analysis

Upgrade via patch package: 8.2.2-beta-patch20260626

Compatible with Lewei Operations Agent v8.2.1 and above

Whether enterprises require highly compliant on-premises private deployment, cost-efficient cloud-hosted solutions, or an AI workbench for batch analysis and operational knowledge accumulation, a complete closed-loop full-toolchain operations agent is now ready for use.

AI should never be merely a chat tool for operations personnel. Instead, it serves as a dedicated operations partner on duty 24/7, familiar with business logic, capable of troubleshooting, executing self-healing and retaining accumulated operational expertise. On late nights hit by waves of alarms, it will shoulder most repetitive and trivial workloads, freeing operations staff to focus on architecture optimization and business innovation rather than endless fire-fighting.

Moving forward, we will keep iterating more industry-specific AI scenarios and optimizing self-healing execution capabilities. Together with our users, we will advance toward the next generation of fully intelligent, lightweight, closed-loop IT operations.