Agentic Observability: How Cloud Monitoring 2.0 Lets Agents Understand Systems and Themselves

As AI agents become the primary operators of software systems, observability must evolve from human-centric dashboards to machine-consumable fact systems. Alibaba Cloud's Cloud Monitoring 2.0 addresses this with Agentic Observability, comprising two symmetrical capabilities: Observability by Agents and Observability of Agents.

Observability by Agents: From Display System to Fact Supply System

IDC predicts active agents will grow 40x to over 1 billion by 2029, with agent API/token load increasing 1,000x by 2027. Since agents — not humans — now read telemetry, the platform must deliver complete, clean, semantically rich, and task-scoped facts via machine-friendly interfaces.

Complete ingestion across full stack and multi-cloud

Cloud Monitoring 2.0 now supports 500+ cloud services and components, a 150% increase year-over-year, covering major providers (AWS, Azure, Tencent Cloud, Volcano Engine). Ingestion spans AI agent runtimes (OpenClaw, Hermes, Codex, Qoder, Claude Code), user endpoints (Web/H5, iOS, Android, mini-programs, Electron), microservices (Java, Go, Python, Node.js, .NET, OBI), self-hosted middleware (MySQL, MongoDB, ClickHouse, Kafka, Redis), infrastructure (hosts, containers, GPU, Sandbox), and open standards (OpenTelemetry, Prometheus). Data types include metrics, logs, traces, events, entities, and profiles.

Intent-based intelligent onboarding

Onboarding shifts from form-filling to intent-driven: the agent describes the goal; the system discovers cloud accounts and data sources; generates a plan with matched components and parameters; presents a preview showing scope and changes before deployment; then deploys and validates with real samples. This makes changes visible before they happen and ensures validation uses actual data, not just configuration success.

Data governance pipeline (Pipline) for actionable telemetry

Raw telemetry is noisy and semantically fragmented. The new Pipline provides three stages: (1) Unified ingestion endpoint accepting OpenTelemetry and Prometheus; (2) Governance & enrichment standardizing fields, enriching tags, and establishing cross-source relationships; (3) Cost routing applying filtering, sampling, aggregation, tiering, routing, and on-demand storage to reduce noise and cost. All four signal types (metrics, traces, logs, events) flow through this single pipeline, producing Actionable Telemetry consumable by both humans and agents.

UModel: unified semantic layer for agent understanding

UModel organizes dispersed signals into a semantic network across five dimensions: entity semantics (unique identity), data semantics (meaning of each metric), relation semantics (call, run, dependency topology), knowledge semantics (domain know-how), and operation semantics (runbooks as executable actions). The platform now ships 3,300+ built-in UModel definitions across 10 observation domains and 214 model groups (85% growth). Custom modeling lets teams define business entities, data, and relations, with one-click publishing. Integrated with STAROps, AI-assisted modeling auto-discovers entities, infers relations, pre-fills semantics, and generates models for human confirmation.

Three agent-native consumption entry points

Skill (method layer) : encapsulates domain knowledge and standard workflows, telling agents how to perform observability tasks.

CLI / CMS2 CLI (execution layer) : deterministic local command invocation for scripting, automation, and auditability.

Cloud MCP Server / CMS Cloud MCP (connection layer) : cloud-hosted standard MCP endpoints for native agent integration.

Together they expose 200+ tools across 21 primary domains, covering 95% of Cloud Monitoring 2.0 functions, returning structured Text/JSON. The roadmap targets 100% coverage.

Observability of Agents: Making Agent Behavior Explainable, Drillable, Governable

Agent systems are non-deterministic: a green trace does not guarantee correct output; token costs and retries scatter across layers; root causes may reside in inference services or GPU infrastructure far below the agent layer.

Five-layer full-stack coverage

Cloud Monitoring 2.0 observes AI User End, AI Runtime, AI Inference Service, PaaS Platform, and AI Infra. Each layer contributes dedicated ingestion and core metrics.

Four capabilities chained on a single trace

User scene replay : RUM-based Session Replay reconstructs user waits, client stalls, stream interruptions; Web Vitals and profiles pinpoint interaction performance; JS errors, blank screens, crashes correlate with device and network context.

End-to-end trajectory and full-stack topology : restores LLM, Skill, Tool, MCP, Memory calls aggregated by User/Session/Trace; correlates metrics and logs; supports drill-down from any hop to full-stack topology — multi-agent calls down to Sandbox, AI Gateway, inference service, GPU, and up to terminal and database.

Multi-dimensional analysis dashboards : accumulates token usage, cost, and model performance along the chain; real-time token cost distribution; comprehensive model call metrics; out-of-the-box dashboards for tools, skills, memory, RAG.

Intelligent clustering and attribution : semantically clusters traces/sessions; attributes performance, cost, quality, security issues; combines trace anomaly diagnosis, session trajectory analysis, and agent health inspection to surface problems automatically.

Closed Loop: Agents Understand Systems, Systems Understand Agents

Both halves share the Cloud Monitoring 2.0 foundation. Agents consume facts to improve system operations; the platform observes agents to improve agent reliability. Each cycle strengthens both sides. As IDC forecasts a 1,000x surge in system access frequency, observability becomes the trusted foundation between humans and agents — humans trust it to delegate authority; agents trust it to make decisions. Cloud Monitoring 2.0's Agentic Observability aims to lay that foundation.