Why Tencent Music Rejects AI Hype: Building an OpenClaw‑Powered Intelligent Ops Ecosystem

Evolution of Tencent Music’s Intelligent Operations

Based on Bian Xuedong’s presentation at the 2026 XCOPS Intelligent Operations Management Conference, the company has iterated a fully covered, closed‑loop digital ops system that consists of three tightly coupled layers: a DevOps delivery layer, a middle‑tier SRE operations layer, and a cloud‑native foundation layer built on Kubernetes.

Traditional Pain Points and Early Optimizations

In 2014, alert overload and manual on‑call caused up to 3,000 monthly alarm calls per engineer. Fixed‑threshold alerts generated fragmented noise, and diverse business workloads required disparate thresholds, leading to missed or false alarms. To address this, a 3‑sigma variance algorithm was introduced (2014‑2015), converting real‑time metric fluctuations into dynamic thresholds based on the last 30 minutes, yesterday, and the same weekday last week. This reduced monthly alerts from >3,000 to around 200.

Advanced Alert Correlation and Root‑Cause Analysis

After the initial reduction, the remaining challenge was fragmented alerts from a single fault. A custom alert workflow was built that automatically links related alarms using document retrieval and semantic extraction, and a toolbox integrating observability, component anomaly detection, K8s monitoring, and full‑chain DevOps data was added. Dify‑based standardised workflows further improved intelligent correlation and anomaly detection.

Full‑Link Observability and Precise Fault Attribution

The observability stack now aggregates node‑level metrics, IP probing, and network jitter, enabling single‑point fault localisation that is then chained to full‑link alerts mapped to business‑level SLA indicators. Micro‑service reporting windows are tuned (core services 20 s, others 1 min) to achieve second‑level anomaly detection. Code‑level tracing links failures back to specific commits and functions.

Massive Data Architecture

To handle up to 4 billion data points per minute, a three‑tier data pipeline was built: a fault‑tolerant ingestion layer, a storage layer using a compute‑separate architecture with StarRocks (migrating from Elasticsearch), and an application layer feeding Grafana, Superset, and alert channels (WeChat, phone). A custom DC Agent captures fine‑grained music‑specific metrics such as playback errors.

AIOps Capability Upgrade

Beyond the mature ops foundation, AI was introduced to empower data‑driven alert optimisation. Scenario‑aware tags (e.g., holidays, promotions) enable dynamic alert throttling. The AI‑enhanced AIOps stack is defined as a "cloud‑native base + AI‑intelligent ops" two‑wing architecture, with fault‑self‑healing currently operating under an "AI analysis + human confirmation + automated execution" safety model.

OpenClaw: A Secure, Multi‑Agent AI Ops Framework

OpenClaw ("Lobster") was launched to provide a sandboxed, fully isolated environment on top of the existing K8s platform. Input‑layer security intercepts and validates all traffic via a unified gateway and proxy; output‑layer security encrypts large‑model and external requests in collaboration with Tencent Zhuque Lab. A dedicated high‑security OpenClaw cluster safeguards code and development data.

The system adopts a "master Agent + sub‑Agents" model: the master Agent schedules tasks, while sub‑Agents embody role‑specific capabilities (architect, product, ops, dev). This design isolates permissions, supports multi‑scenario analysis, and avoids single‑agent limitations.

Unified Agent Runtime and Core Services

To avoid redundant integrations, a generic Agent Runtime abstracts various AI models (Hermes, multimodal, OCR, code generation) behind a unified gateway, skill marketplace, and plugin service. Two core services were built:

Super Memory : a knowledge‑graph‑enhanced long‑term memory that reduces per‑conversation token usage from 50‑60 k to 10‑20 k, mitigating context explosion.

Super LLM : an intelligent router that selects low‑cost internal models for simple Q&A and high‑performance models for code‑heavy tasks, while throttling unnecessary token consumption.

Implementation Philosophy and Future Roadmap

The team stresses "scenario‑first, problem‑solving" over blind AI stacking. Currently, AI is confined to deterministic fault‑analysis scenarios, with any production actions gated behind a one‑click human approval. Future work will deepen AI analysis, enrich the knowledge base, and expand multi‑Agent coordination to move from assisted to autonomous operations.