AI Innovation Forum Shanghai: Key Takeaways, Multi‑Agent Architecture, and PPT Resources

The Cloud Native team, together with the Public Cloud Shanghai regional CTO line, co‑hosted the AI Innovation Practice Forum, which attracted more than 70 technical participants.

Agenda 1 – Accelerating Enterprise Agent Engineering

Speaker Meng Fang (Alibaba Cloud Intelligent Cloud‑Native Application Platform) introduced the “Agent Native Cloud” three‑layer capability model: an Infra layer providing runtime foundations, a Platform layer addressing scale through high‑code AgentRun construction, no‑code AgentTeams governance, and AgentLoop observability optimization, and a Desktop layer connecting to business. The session also showcased the STAROps full‑stack intelligent operations platform for ensuring stability of emerging agent services.

Agenda 2 – AgentTeams: One‑Stop Agent Harness

Speaker Cui Feifei (Alibaba Cloud Intelligent Senior Solution Architect) described AgentTeams as an enterprise‑grade multi‑agent collaboration and governance platform. It adopts a Manager‑Team Leader‑Worker hierarchical architecture, leverages a Matrix‑protocol transparent group chat for human‑machine and agent‑agent interaction, and enforces four security layers via zero‑trust gateway and sandbox isolation. The platform unifies heterogeneous agents such as QwenPaw, OpenClaw, and Claude Code.

Agenda 3 – AgentLoop: Full‑Lifecycle Evolution Flywheel

Speaker Xia Ming (Alibaba Cloud Intelligent Senior Product Expert) explained that AgentLoop tackles four production challenges—quality, cost, performance, and security—by building a data‑flywheel centered on “Trajectory” as the backbone. The closed loop consists of observation & audit → evaluation & experiment → continuous optimization. Features include non‑intrusive end‑to‑end data collection, full‑link behavior audit, an “Agent‑as‑a‑Judge” evaluation paradigm, and CI/CD experiment gating, enabling precise effect optimization and risk control.

Agenda 4 – Building the Network Foundation for the Agentic AI Era

Speaker Jin Zhan (Alibaba Cloud Intelligent Solution Architect) addressed rapid start‑stop, massive concurrency, and security control challenges of large‑scale AI agents. He presented a seven‑layer cloud network solution comprising VPC, NAT, ALB, EPG (outbound proxy gateway), and GA (global acceleration). The network shifts from a passive “connection pipe” to an active “traffic‑decision layer,” continuously handling entry, permission, and path decisions to provide high‑performance, secure, global infrastructure for agents.

Agenda 5 – Large Model + Observability Graph: Next‑Gen AIOps Practice

Speaker Cao Rui (Alibaba Cloud Intelligent Solution Architect) introduced STAROps, which upgrades traditional AIOps to “Agentic Ops” for 24/7 autonomous operations. Built on a unified observability data foundation and a digital twin (UModel) that captures system context, STAROps offers an intelligent assistant (natural‑language diagnostic), long‑running automated inspection tasks, and a digital employee (dedicated SRE). These capabilities transform operations from reactive firefighting to proactive platform management, markedly improving system stability and efficiency.

Closing

The event concluded with on‑site photos and a downloadable collection of speaker PPTs for further reference.