Alibaba Cloud's Agent-Friendly Computing: From 'Computable' to 'Executable' at Yunqi 2026

Paradigm Shift: From Model Era to Agent Era

Alibaba Cloud's senior product expert Ji Shaochen highlighted that the core demand has shifted from "computable" (model era) to "executable" (Agent era). Agents must maintain context, re-plan across multi-step reasoning, frequently call tools (browsers, code environments, databases, enterprise systems), and require both high throughput and low latency, with long-running tasks and bursty fine-grained workloads.

This shifts infrastructure needs beyond GPU inference: data preprocessing, code execution, tool calling, task orchestration, RL rollouts, and runtime state management run heavily on CPU. General-purpose compute becomes the critical foundation for stable Agent deployment.

Three-Layer Architecture for Agent-Friendly Computing

Layer 1: Agent-Friendly Core Technologies

Innovations in high-density deployment, image pre-warming, hibernation/recovery, and VM optimizations handle fine-grained, ultra-concurrent tasks, enabling rapid Agent startup, efficient idle hibernation, and state preservation for long tasks.

Layer 2: Agent-Friendly Compute Products

Products tailored for Computer Use, Mobile Use, coding agents, browser agents, data agents, and industry agents. Offers x86/ARM, containers, native VMs. Optimized for Agent memory footprint, tool calls, network access, isolation, and startup speed — not just average generic workloads.

Layer 3: Agent-Friendly User Interface

CLI, MCP, Skills let Agents understand, call, and orchestrate cloud resources. Risk confirmation, cost estimation, human authorization, result verification, and full audit trails turn natural language into a secure, controllable cloud interface.

Four New Products: Two Groups

Group 1: Agent on ECS (Optimizing Agent Runtime)

Agent-Specific Instance

Packages AI app, cloud server, and model inference service. Starting memory +50%, fixed public IPv4, peak bandwidth, model quota. Pre-installed Agent app images; integrates Qwen, Wanxiang, DeepSeek, Zhipu, Kimi, MiniMax. Agentic OS reduces invalid token consumption 30%, boosts benchmark scores 10%, memory allocation efficiency 200%, cold start -20%. Entry config: 2 vCPU, 6 GB RAM, 30 GB disk, fixed IPv4, model credits ~¥2.9/day — ~20% cheaper than "ECS + MaaS" combo, 50% more memory, 25% more tokens.

10th-Gen Enterprise Instance g10a (Invite-Only Preview)

AMD next-gen CPU, memory channels 12→16 (+60% bandwidth). Alibaba Cloud Linux 4 OS-level Agent optimizations. CIPU boosts cloud disk/NIC density. vs 9th-gen: ~30% perf gain in Agent coding, ~50% in search/recommendation, ~25% better price-performance. Full-stack upgrade for Agent workflows, data, toolchains.

Group 2: Agent in ECS (Platform Drives Cloud Operations)

E-HPC Agent

For HPC: user defines goal/boundaries; Agent understands task, plans experiments, calls industry Skills, iterates based on results. Expert handles key judgments; Agent automates tool discovery, pipeline stitching, resource scheduling, result analysis. Closed loop: goal understanding → experiment planning → execution feedback.

Case: Shanghai Institute for Scientific Intelligence's "Huntianling" platform (140+ Skills). 5-day wet-dry lab loop: 867 compute tasks, 70k+ CPU core-hours, 829 hrs runtime. 102 planned tasks done in ~2.5 days; 30+ added. Catalyst screening: 2nd-round conversion rate 6x original. Researchers focus on questions/judgments; Agent + elastic compute handle repetitive execution.

Elastic Computing Agent

Covers resource purchase, management, ops full chain. Not a chatbot: perceives issues, analyzes root cause, executes ops, verifies results. Accessible via console, DingTalk, Feishu, WeCom, OpenAPI, Skills. Main Agent + specialized sub-Agents + scenario Skills cover knowledge Q&A, purchase assist, resource mgmt, app monitoring, diagnosis/optimization. Generates cost/performance/price-prioritized purchase plans; natural language for instance/disk/security group ops and IaC deployment. GPU cluster inspection: fault detection/localization hours→minutes, >80% manual effort saved. Human retains control: high-risk ops need confirmation, full audit trail. Division: human = goal + authorization; Agent = planning + execution + feedback.

Cloud Assistant OSAgent

Brings AI judgment into OS layer. Uses Cloud Assistant's native secure channel to build perceive-plan-act-feedback loop in Guest OS. MCP for OS metadata; Skills reuse expert knowledge. Adapts to Ubuntu/CentOS/Red Hat/Windows per instance, collapsing multi-instance, multi-OS, multi-command workflows from dozens of interactions to one conversation — ~10x ops efficiency. For complex issues (kernel panics), correlates live logs, core dumps, kernel source, community changes to pinpoint root cause and repair within approval scope. Approval, whitelist, least privilege, logging, rollback, audit throughout. OSAgent and Skills open-sourced. Cloud Assistant handles 44M+ tasks/day, 99.9%+ success rate.

Real-World Validations

StepFun (Harness Platform)

Post-training & RL platform on ECS g9i (Intel Xeon 6 + CIPU). Control-plane optimization, image pre-warm, ACS K8s compatibility, dynamic ephemeral storage. Stable at hundreds of thousands of cores. g9i: +20% perf, -5% unit compute cost. Harness perspective: infra offloads context mgmt, model-cost matching, interruption recovery, multi-Agent orchestration — developers focus on goals, acceptance criteria, baselines, release gates.

Haier Smart Home

Global 100+ countries, 100M+ registered users, 10M+ MAU. ECS g9i high-perf compute, network forwarding, elastic scaling supports massive concurrency, ms-level device interaction, s-level voice control. g9i provides stable, reliable compute foundation for existing scale, enabling upper-layer product innovation.

9th-Gen AMD Instances

CIPU architecture, 400G network, 1 vCPU = 1 physical core, 384 cores/server. Three scenarios:

Search/Recommendation/Ads: large-model embeddings, tokenization, feature extraction, sparse data → ~37% perf gain in customer benchmarks; million-core clusters deployed.

Small/Mid Voice Models: 7B/14B models viable for verticals. Low latency critical; high core density enables multi-tenant concurrency, commercial viability.

Agent Runtime & RL Rollout: code exec, tool calls, browser ops, state mgmt, orchestration on CPU; Rollout needs massive concurrent interaction paths. Physical cores, compute, memory bandwidth boost throughput. CPU becoming new system-level bottleneck for Agent efficiency.

Yizhi Intelligence (Voice Agent)

Voice Agent ≠ simple ASR+LLM+TTS chain. Real-time turn-taking judgment, interruption handling, balance intelligence/latency/stability/cost. Turn-taking module + Fast Thinker routes simple queries to fast models, complex to stronger models/background chains. Heterogeneous deployment: Agent Runtime, state mgmt, full-duplex control, LLM gen, TTS on matched CPU, general compute, GPU. Auto-scaling for traffic peaks/valleys → 40% compute cost reduction. Key: match tasks to right compute, not just pile hardware.

Conclusion

"Agent-friendly" means systemic rebuild across core tech, compute products, and user interface — not just a chat entry or more compute. Bottom layer handles fine-grained, high-concurrency, long tasks; middle matches diverse Agent workloads; top is understandable, callable, orchestratable by Agents with human authorization, risk control, verification in the loop. As Agents scale across industries and become cloud users themselves, compute services evolve from resource providers to intelligent execution substrates. Alibaba Cloud Elastic Compute will continue evolving around elasticity, cost, stability, security — making Agents runnable, stable, affordable, controllable — solid compute backbone for every enterprise's AI evolution.