How Manbang Cut Root-Cause Analysis to 1 Minute with AgentCore-Powered SRE Digital Human
Manbang leverages Alibaba Cloud AgentCore to build a production-grade SRE digital human that integrates business knowledge, tools, and runbooks, achieving a median root-cause analysis time of 1 minute with over 90% accuracy across core freight-platform workflows.
Customer Background: Stability Demands Behind Digital Freight
Manbang is a digital freight platform connecting shippers and drivers. Stable operation of cargo sourcing, shipping, and order business chains is critical. A single order or shipping anomaly may involve multiple applications and middleware, requiring timely business impact assessment, root cause identification, and remediation.
SRE Challenges: Cross-Tool Collaboration and Complex Business Chain Diagnosis
Manbang has built a stability product matrix covering prevention, detection, response, recovery, and review, including the Lancet governance platform, Hubble, and tools for on-call, inspection, and drills. However, a typical on-call engineer must monitor 5 systems and open 6 platforms during an incident. Information needed for a single investigation is scattered across different systems, requiring manual correlation and judgment. Root cause analysis is further complicated by long, complex business chains (shipping, orders) where alert locations may not be the true source. Multiple simultaneous metric anomalies require understanding which applications and interfaces correspond to which business, their service dependencies, and the impact scope.
When intelligent agents join the workflow, production requirements emerge: conclusions must be evidence-based, tool calls authorized, operations traceable, with human intervention and rollback mechanisms. Connecting tools, understanding business, and constraining execution must be considered in a single workflow.
Solution Practice: Connecting Tools, Knowledge, and Diagnostic Processes with AgentCore
Manbang first mapped existing professional capabilities (inspection, on-call, root cause analysis) to AgentCore components. Alibaba Cloud AgentCore is an enterprise-grade agent building and governance platform providing unified support for construction and operation, collaboration and governance, observation and evaluation. Enterprises can create and manage Agents in a unified console, connect different models, integrate external tools via MCP, encapsulate and reuse professional capabilities via Skills, and reduce repetitive configuration with Agent templates. AgentCore also provides team collaboration, role-based permissions, credential management, runtime monitoring, and integration with DingTalk and Feishu for daily communication channels.
This division of labor means Manbang's existing stability systems, business knowledge, and diagnostic experience continue to function, while AgentCore handles generic runtime and governance capabilities. The team focuses on professional scenarios like inspection, root cause analysis, and controlled remediation.
Cross-Tool Collaboration via MCP and Skills
Manbang uses MCP to integrate professional capabilities through tool calls, reuses standard operating steps via Skills, and orchestrates multiple specialized agents to collaborate on knowledge matching, call-chain evidence collection, and comprehensive analysis. AgentCore's managed runtime, isolated environments, and elastic scaling support tasks so that tools can participate in a single investigation rather than merely outputting information separately.
Real-Time Inspection Combined with Regular Governance
Automated inspection runs every 10 minutes: if no anomaly is found, current metric levels are pushed via DingTalk; if anomalies are detected, risk cards, diagnostic results, and remediation suggestions are pushed. The system leverages Hubble call-chain analysis, code scanning, and rule engines to identify cross-application risks. Discovered risks enter the Lancet platform for governance, and completed governance is verified through drills, closing the loop of risk discovery, remediation, and acceptance.
Business Understanding via Knowledge Graph
AgentCore provides memory, knowledge base access, and context recall. Manbang builds business knowledge by associating application, interface, middleware, and infrastructure archives with stability rules and business relationships. Data is continuously collected, cleaned, and refined to interface granularity, enabling the digital human to first determine the business context of an anomaly and then narrow the investigation scope along the knowledge graph. Business teams maintain their own data, which also serves as the basis for inspection and diagnosis of their services.
In this process, basic data comes from actual collection; deterministic metric judgments are executed by code and rules; Agents handle task scheduling, evidence-based analysis, and summarization. Collection and judgment processes are traceable and auditable, facilitating on-call engineer verification of the facts underlying model analysis.
Production Control: Permissions, Approvals, and Rollback
AgentCore enforces production guardrails through a built-in AI gateway, Agent identity, and policy control, establishing permission boundaries for access and invocation while providing audit and process tracing. Manbang incorporates approvals, runbooks, execution records, and rollback mechanisms into the remediation workflow. The platform manages tool access; business processes constrain production actions; diagnostic suggestions do not equate to execution authorization.
On this foundation, Manbang is advancing dual-track collaboration between root cause analysis and application self-protection: Agents trace and locate issues and provide recommendations, while application-side SDKs monitor related metrics at second-level granularity, and a decision center drives self-protection according to preset strategies. The self-recovery scope is explicitly limited to four preset actions—throttling, degradation, scaling, and isolation—executed under authorization and runbook constraints. Runbooks enter the execution library only after drill verification, and execution retains records and rollback constraints.
Practice Results: Median Root Cause Location Time of 1 Minute
According to Manbang's practice data, root cause analysis now covers core chains with a median location time of 1 minute and accuracy greater than 90%. In core scenarios, the digital human can first clarify business ownership, narrow the investigation scope, and then perform call-chain drill-down and root cause analysis; capabilities continue to improve.
For example, with user feedback like "order placement failure" or "shipping failure," on-call engineers first need to identify the involved business segment and applications. The digital human identifies the business scenario and domain from the description, maps to applications and interfaces, then drills down along relevant call chains to collect evidence, incorporating call relationships, changes, and metrics into the analysis. If multiple feedback items appear simultaneously, they are first categorized and analyzed for commonalities to distinguish client-side, server-side, or other issues.
After evidence collection, the model synthesizes a diagnostic result. On-call engineers receive associated objects, diagnostic evidence, and suggested actions for verification before proceeding with remediation per the authorization process. Investigation converges from concrete business problems to relevant objects without initially traversing all applications; analytical conclusions are backed by evidence for review and sharing. This demonstrates the digital human's role: organizing scattered information, providing a basis for human judgment, and pointing the investigation direction.
Automated inspection pushes risk information to daily collaboration channels; on-demand analysis lets developers investigate as needed; evidence-backed conclusions facilitate team review. The digital human's application thus integrates into Manbang's existing on-call, diagnosis, and governance processes, with professional staff retaining judgment and remediation responsibility.
Future Directions: Expanding Diagnostic Coverage and Advancing Controlled Self-Recovery
Building on core scenario adoption, Manbang will gradually expand business coverage, further connect knowledge graphs across client-side, traffic entry, gateway, and application layers, and enrich interface functions, application tiers, business impact, and expert remediation experience. In analysis capabilities, the team will continue exploring investigation scope trimming, anomaly scoring, and change-information weighting. Self-recovery will advance by expanding the library of verified runbooks within clearly authorized scopes.
AgentCore's integrated AgentLoop observation and evaluation capabilities support tracing model invocations, tool usage, and session processes. Manbang plans to combine this capability to feed back verified root cause conclusions, review results, and failure samples as reusable data and Skill assets, adjusting processes for knowledge gaps and inaccurate judgments exposed in practice, and validating improvements through evaluation and experimentation.
Manbang aims for the stability digital human to gradually assume responsibilities at the stability lead level. Achieving this requires continuous capability refinement and validation in more business scenarios, with diagnostic capability improvement and execution scope expansion verified separately.
Manbang's practice offers a construction path starting from the existing stability system: business knowledge directs investigation, actual observations support diagnosis, and authorization and runbooks constrain remediation. Manbang's professional accumulation and Alibaba Cloud AgentCore's platform capabilities each have distinct roles, collaborating in concrete tasks to enable enterprises to gradually expand the scope of agent participation in production work based on existing tools and experience.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
