Alibaba Cloud STAROps Upgrade Introduces Agentic Ops for Closed-Loop Problem Resolution

After an application release, error rates spike. Alerts fire, and monitoring, logs, and traces are available, yet engineers must still switch across systems to confirm impact scope, trace upstream and downstream dependencies, correlate recent releases, and decide on remediation. Even when AI provides analysis suggestions, the work is not done: engineers must verify the evidence behind the suggestion, assess operational safety, and confirm business recovery after execution. This gap — data and AI make detection easier, but the path from detection to resolution remains heavily manual — is the common challenge enterprises face when adopting intelligent operations.

From AIOps to Agentic Ops

Rule-based monitoring helps detect anomalies; AIOps adds anomaly detection and correlation. With large language models entering operations, users can now query data, explain alerts, and analyze problems in natural language, lowering the usage barrier. However, in many scenarios, deciding what to query next, when to continue, and whether remediation restored service still requires human drive.

Agentic Ops changes this workflow: users set a continuous operations goal, and an Agent observes the environment, detects anomalies, investigates, and drives remediation and verification within authorized boundaries. For example, a user can ask the system to continuously monitor a core application's releases, check runtime status after each new version, correlate changes when anomalies appear, and follow up on subsequent handling. The core shift is from humans repeatedly asking and pushing, to Agents working continuously around a goal while humans make key decisions.

Unified Operations Context: Letting Agents Understand the Real Environment

Enterprise operations data is scattered: metrics in monitoring platforms, runtime logs in log platforms, assets in resource systems, code and release changes in development platforms. These together form the basis for investigation. STAROps uses automatic perception, intelligent onboarding, unified connection, and data modeling to organize them into an operations context that Agents can understand and analyze.

Automatic Perception and Intelligent Onboarding, Lowering Integration Barriers

Within authorized scope, STAROps automatically discovers cloud resources and continuously senses resource changes, reducing manual environment mapping. During data onboarding, users describe their integration intent in natural language; the system matches data sources and onboarding plans against the current environment, assisting configuration and validation so users reach analysis and diagnosis faster.

Connecting Multi-Source Systems, Unified Data Modeling

Connectors and Fabric share the work of building a unified context:

Connector (Unified Connection): Via direct cloud connections and Satellite environment proxies, connects cloud environments and enterprise systems, centrally managing authentication, resource discovery, health checks, and connection status. Coverage spans observability data, resources and costs, and extends to code, CI/CD, ticketing, and automation systems.

Fabric (Unified Modeling): Associates metrics, logs, traces, and other observability data with application, instance, container, database, and other resource objects, continuously maintaining unified semantics, entity relationships, and dynamic topology.

When an interface errors, the Agent needs to know which service it belongs to, which instances run it, which database it depends on, and whether recent changes are related. Such context enables subsequent cross-system, cross-product investigation.

From Natural Language Query to Evidence-Supported Cross-Domain Diagnosis

With unified context, users can start from a simple question:

Help me check APM applications with high error rates in the last hour.

STAROps understands the query object and time range, queries relevant data, filters, sorts, compares, and extracts conclusions needing attention. Users can continue asking: which interfaces are mainly involved? What errors appeared? Are there anomalies in associated traces? Subsequent analysis reuses existing context, reducing repeated background description and system switching.

Q&A is one entry point; the same data and analysis capabilities can be invoked by alert events, long-running tasks, and problem discovery. When the question deepens to "why did the anomaly occur," a diagnosis process backed by evidence is needed. STAROps first locks the anomalous object, time window, and impact scope, then combines resource dependencies to correlate metrics, logs, traces, changes, code, and other evidence. On this basis it forms hypotheses, continues evidence gathering and verification, and progressively converges on the cause. For instance, a release just before the anomaly becomes a lead, but affected instances, error stacks, and call chains must verify the judgment; if evidence is insufficient, investigation continues or the analysis direction adjusts. The final diagnosis must state root-cause judgment, business impact, propagation path, supporting evidence, and remediation suggestions, helping users understand the basis for the conclusion and decide next steps. This process applies to cross-cloud-product failures, application end-to-end diagnosis, change anomaly localization, and focused investigation among large alert volumes.

Release, Resolve, Verify: Continuous Closed-Loop, Safe and Controlled

After one diagnosis, the problem must still be tracked. STAROps connects continuous observation with problem handling through long-term tasks (Missions) and problem discoveries (Findings).

Long-Term Tasks (Missions): Continuous Operation Around Operations Goals

After users define the business scope and strategy, Missions continuously observe system changes, detect anomalies, investigate, and generate Findings when remediation tasks are discovered. Goals can be cross-cloud resource patrols or continuous attention to an application's release impact. Task execution does not depend on users repeatedly initiating new queries.

Problem Discoveries (Findings): Carrying the Full Problem-Solving Process

Phenomena, evidence, impact, root-cause judgment, and remediation suggestions from investigation are aggregated into a continuously updated problem object. Subsequent ownership, progress, executed actions, and final results are recorded in the same problem. Even across multiple investigation rounds or human interventions, the same problem context persists for continuous advancement and resolution.

Experience Accumulation: Every Remediation Becomes Experience

After execution, actual effect must be verified. If business has not recovered, the problem continues investigation or escalates to human handling. Tool-call success alone does not prove resolution. Investigation processes, remediation trajectories, and verification results are continuously accumulated for reference and reuse by subsequent tasks.

Tiered Execution: Defining Production Operation Boundaries

Before production execution, STAROps forms a remediation plan specifying target objects, concrete actions, and impact scope, then advances per execution strategy:

Auto-execute: Low-risk queries, analysis, reporting allowed by permissions and policies.

Human confirmation: Configuration changes, restarts, scaling, and other critical actions.

Hard block: Operations hitting deny rules.

Whether auto-approved or human-confirmed, policy and target-resource permissions are re-validated before execution; human confirmation cannot bypass existing permission boundaries. Post-execution recovery is verified; data access, tool calls, approvals, and execution results are logged. If expectations are not met, the system rolls back per preset plan or escalates to human handling.

Multiple Integration Modes, Fitting Into Enterprise Operations Workflows

Enterprises already have their own operations platforms, development tools, and collaboration methods. STAROps offers multiple integration modes so capabilities enter existing flows: OpenAPI for query, analysis, and diagnosis; embedded page components in existing workbenches; collaboration entry points in Feishu, DingTalk, and WeCom for data query and diagnosis collaboration; and Skill, CLI, Plugin interfaces for teams using AI tools or self-built Agents. Unified capabilities and context serve different user entry points, letting users complete tasks in familiar environments.

Landing intelligent operations requires a continuous work process connecting data, judgment, and action. STAROps reduces context fragmentation through unified connection and modeling; helps experts form judgments faster through cross-domain evidence gathering and diagnosis; and continuously drives problem resolution through long-term missions, tiered execution, and effect verification. The aim is to let Agents shoulder more continuous operations work while humans focus on key decisions and complex problems, enabling a fundamental shift from reactive response to proactive autonomy and providing 7x24 autonomous operations capability that guards business system resilience in real time.