Alibaba Cloud Agent Native Cloud: AgentRun, AgentLoop & AgentTeams Guide (2026)

On July 18, 2026, Qi Zhou — head of Alibaba Cloud’s Cloud-Native Application Platform — walked onto the stage at the World Artificial Intelligence Conference (WAIC) in Shanghai and made an announcement that most Western enterprise architects will take weeks to fully process: Alibaba Cloud is no longer just running AI workloads on cloud infrastructure. It is rebuilding the cloud itself around AI agents as the primary compute primitive.

The announcement introduced Agent Native Cloud — a suite of three coordinated services (AgentRun, AgentLoop, and AgentTeams) plus a new inference engine (TokenWorks) and an open-sourced chip software stack (T-Head SAIL). Together, they represent the most vertically integrated agent infrastructure stack any cloud provider has publicly assembled: from custom silicon at the bottom to multi-agent governance at the top.

For the rpabotsworld.com audience — RPA developers, solution architects, and automation engineers moving into the agentic AI space — this matters regardless of whether you ever deploy on Alibaba Cloud. The announcement confirms that agent operations is now a sellable cloud primitive, on the same level as compute, storage, and networking. Every hyperscaler is converging on this layer, and the architectural patterns Alibaba is productizing are the same ones you’ll need to understand whether you’re building on AWS, Azure, Google Cloud, or an open-source stack.

This guide breaks down what Alibaba actually announced (and what it conspicuously didn’t), how the stack compares to AWS Bedrock AgentCore, Microsoft Agent 365, and Google’s Agentic Data Cloud, and what practitioners should do about it right now.

What Is Agent Native Cloud? The Three-Layer Architecture

Agent Native Cloud is not a single product — it is a vertically integrated platform comprising three services that address distinct phases of the AI agent lifecycle. Understanding the boundaries between them is critical, because most secondary coverage collapses all three into a single “Alibaba launches agent platform” headline that obscures what’s actually new.

Here’s the precise breakdown, drawn from Alibaba Cloud’s official announcement:

AgentRun: Lifecycle Management (Pre-Existing, Expanded)

AgentRun is not new. It predates WAIC 2026 and provides lifecycle management covering development, deployment, and operations for AI agents. Think of it as the runtime layer — the managed environment where your agent actually executes. It provides:

• Native sandbox environments — isolated execution spaces with strong workload isolation, so one agent’s failure or misbehavior doesn’t cascade
• Elastic scaling — automatic resource allocation based on agent workload demands
• Enterprise identity integration — agents inherit organizational identity and access policies rather than running with their own credentials
• Development-to-production pipeline — a managed path from agent prototyping through staging to production deployment

If you’ve worked with container orchestration platforms like Kubernetes, the mental model is similar: AgentRun is to AI agents what a managed Kubernetes service is to containerized applications — the runtime layer that handles the operational mechanics so you can focus on the agent logic itself.

More Read

Working with UiPath Test Automation

RPA Center of Excellence: Building and Sustaining Success

Telegram Bot Integration With UiPath (Remote Captcha solving)

AgentLoop: Observability and Optimization (New at WAIC 2026)

AgentLoop is the first of two genuinely new services announced at WAIC. It provides real-time tracing, evaluation, and optimization of agent performance. In practical terms, this is the agent-equivalent of an Application Performance Monitoring (APM) suite — the tooling that most teams currently assemble from a patchwork of tracing SDKs, eval harnesses, and prompt dashboards.

What AgentLoop centralizes:

• Real-time tracing — following the complete execution path of an agent through tool calls, LLM invocations, memory reads/writes, and inter-agent communications
• Performance evaluation — measuring agent accuracy, latency, token consumption, and task completion rates against defined benchmarks
• Optimization recommendations — surfacing actionable improvements based on observed agent behavior patterns

For teams currently building agentic systems with frameworks like LangGraph, CrewAI, or the Microsoft Agent Framework, this maps to the observability gap that everyone acknowledges but few have solved well. Today, most teams instrument their agents with custom logging, run periodic eval suites offline, and debug production issues by reading raw logs. AgentLoop’s promise is to replace that manual stack with a managed service.

AgentTeams: Multi-Agent Governance (New at WAIC 2026)

AgentTeams is the second new service, and arguably the most strategically significant. It provides coordination and governance across multiple agents, allowing organizations to manage complex workflows that involve agent-to-agent collaboration.

This addresses what is rapidly becoming the hardest problem in enterprise agent deployments: once you have more than one agent operating in production, who decides which agent handles which task? How do you prevent two agents from taking conflicting actions on the same data? How do you enforce organizational policies across an entire fleet of autonomous agents?

AgentTeams tackles the agent control plane problem at the infrastructure level:

• Multi-agent routing — directing incoming tasks to the appropriate specialized agent based on capability, context, and policy
• Hand-off protocols — managing the structured transfer of context and control when one agent delegates work to another
• Governance policies — enforcing organizational rules about what agents can do, what data they can access, and what actions require human approval
• Workflow coordination — orchestrating complex multi-step workflows that span multiple agents, with state management and error recovery

If you’re building multi-agent systems today, you know this layer. It’s the part you’re currently implementing with custom orchestration code — the supervisor agents, the routing logic, the shared state stores, the policy enforcement hooks. AgentTeams says: that’s infrastructure, and it should be managed by the cloud provider.

Below the Agent Layer: TokenWorks and T-Head SAIL

The agent services grabbed the headlines, but two lower-in-the-stack announcements may prove more immediately relevant for practitioners.

TokenWorks: Unified Inference Serving

TokenWorks is a new service within Alibaba’s Platform for AI (PAI), available through PAI-EAS. It integrates four functions that are typically handled by separate systems into a single service:

For teams running inference at scale — which includes anyone operating agentic systems where each task may trigger dozens of LLM calls — the cost and reliability implications of unifying these functions are substantial. Inference cost is consistently the #1 operational expense in production agent deployments, and the inefficiencies from running these as separate systems compound quickly.

T-Head SAIL: Open-Source Chip Software Stack

Alibaba’s chip unit, T-Head, open-sourced its SAIL (Software AI Layer) stack at WAIC, providing developers with access to the computing framework optimized for Alibaba’s Zhenwu AI chips. The stack spans operating systems, SDKs, and interfaces, and is designed to be compatible with mainstream AI ecosystems.

The numbers behind the silicon are notable: as of April 2026, cumulative Zhenwu chip shipments reached 560,000 units, supporting over 400 customers across more than 20 industries. That’s a meaningful install base, and open-sourcing the software stack is a clear play to expand the developer ecosystem around Alibaba’s custom silicon — the same strategy that made CUDA dominant for NVIDIA.

The Competitive Landscape: Every Hyperscaler Wants This Layer

Alibaba’s announcement does not exist in isolation. It enters a race that AWS, Microsoft, and Google are all running simultaneously, and understanding the competitive positioning is essential for any architecture decision.

AWS Bedrock AgentCore: The Most Mature Offering

Amazon’s Bedrock AgentCore is the most commercially mature offering in this space, with over 1 million downloads since general availability. The AgentCore harness lets you create and run an agent with just two API calls (CreateHarness and InvokeHarness), with the agent running in its own isolated environment with a filesystem and shell, persistent memory across sessions, and the ability to browse the web, call tools through gateway or MCP, and switch model providers mid-session. AWS also recently added batch evaluations, A/B testing, and Bedrock Guardrails integration for evaluating every agent action against prompt injection, harmful content, and sensitive data policies.

The key advantage AWS has: it’s priced and shipping today. The harness itself is free; you pay for the underlying compute and model inference. That transparency is something Alibaba’s announcement conspicuously lacks.

Microsoft Agent 365: Security-First Approach

Microsoft’s approach, rebranded from Copilot Studio security capabilities to Agent 365, centers on the security and governance dimensions. As of July 1, 2026, Microsoft Copilot Studio agent security capabilities through Defender for Cloud Apps are transitioning to Agent 365 observability logs for licensed tenants. Microsoft has also introduced new certification paths — the Agentic AI Business Solutions Architect (AB-100) and AI Agent Builder Associate (AB-620) — signaling that they see agent operations as a distinct professional discipline.

Google: A2A Protocol and the Interoperability Bet

Google’s approach is arguably the most architecturally ambitious. The Agent2Agent (A2A) protocol reached v1.0 in production at 150 organizations, enabling agents on different platforms to communicate through a standardized protocol. This means a Salesforce agent built on Agentforce can hand off a task to a Google agent running on Vertex AI, which can query a ServiceNow agent for IT asset data — all through A2A without any of the three systems needing to understand each other’s internal architecture.

Google also rebranded Vertex AI to the Gemini Enterprise Agent Platform at Cloud Next 2026, consolidating its agent tooling under a unified brand with 200+ models in the Model Garden.

The Open-Source Alternative: agentgateway

Solo.io donated its “agentgateway” project to the Linux Foundation’s Agentic AI Foundation in June 2026, where it now counts 300+ contributors across 60+ organizations including CoreWeave, Red Hat, Adobe, Salesforce, and Microsoft. This is the open-source counterweight to the hyperscaler lock-in play — a neutral gateway layer for agent-to-tool traffic that doesn’t tie you to any single cloud provider.

The Qwen 3.8 Connection: Why the Model Layer Matters

One day after the Agent Native Cloud unveiling, Alibaba released Qwen 3.8-Max-Preview — a 2.4 trillion-parameter multimodal model with a 1 million-token context window. The timing is not coincidental. The model, the agent infrastructure, and Alibaba’s Token Plan pricing system represent a coordinated strategy: a frontier model (Qwen 3.8), a consumption model (Token Plan), and an operational platform (Agent Native Cloud) — all announced in the same WAIC week.

For practitioners, the relevant question isn’t whether Qwen 3.8 is better than GPT-5.6 or Claude Sonnet 5 (that benchmarking is a separate conversation). It’s that Alibaba is the only vendor that announced a complete vertical stack from silicon to governance in a single event:

No other vendor — not AWS, not Microsoft, not Google — made a stack claim that vertically complete at a single event. Whether Alibaba can execute across all seven layers is an open question, but the ambition sets the direction for what “full-stack agent infrastructure” will mean by 2027.

What Alibaba Didn’t Say: The Critical Gaps

Precision matters in evaluating vendor announcements, and what was not disclosed at WAIC is as important as what was. Across the entire Agent Native Cloud announcement:

• No pricing — for AgentRun, AgentLoop, AgentTeams, or TokenWorks. Not even indicative pricing tiers or a “starting at” figure.
• No general availability date — no preview date, no GA date, no waitlist. The services were “unveiled” and “introduced,” not “launched.”
• No named customers — no pilot deployments, no reference architectures from real deployments, no case studies.
• No executive quote — unusually, no Alibaba executive is directly quoted anywhere in the primary announcement.
• No compliance posture — no mention of SOC 2, ISO 27001, GDPR, or any compliance framework for the agent services.
• No SLAs — no uptime guarantees, no latency commitments, no data retention policies.

This absence doesn’t make the announcement empty — hyperscalers routinely unveil infrastructure directions months before commercial terms follow. But it means the correct classification is stated intent, not a shippable product. Teams should track the direction, not plan migrations around it.

The Gartner counterweight is worth noting here: according to Forbes’ reporting of Gartner’s analysis, more than 40% of agentic AI projects will be canceled by 2027, citing escalating costs, unclear value, and weak risk controls. That prediction is the right lens for any agent infrastructure announcement that arrives without pricing or commercial terms — and it’s a pattern we’ve explored in depth in our analysis of why agentic automation programs fail.

The MCP and Skills Portal: Cloud as Agent Toolkit

One detail from Alibaba’s broader WAIC announcements that flew under the radar of most coverage deserves particular attention from the rpabotsworld.com audience: Alibaba Cloud launched a new Skills portal that converts common cloud capabilities across more than 60 cloud products into Skill-based and MCP-compatible formats.

This is significant. MCP (Model Context Protocol, originally developed by Anthropic) has rapidly become the de facto standard for connecting AI agents to external tools and data sources. Alibaba’s decision to make its cloud services natively MCP-compatible means that an agent running on AgentRun could invoke cloud resources — databases, storage, compute, messaging — as naturally as calling functions, through the same protocol that agents use to call any other tool.

For practitioners working with MCP-aware frameworks (which now includes most major agent frameworks), this lowers the integration barrier significantly. Instead of writing custom API wrappers for each Alibaba Cloud service, you’d use the same MCP protocol you already use for everything else.

Market Context: Why Agent Infrastructure Is the Next Cloud Primitive

The simultaneous convergence of four hyperscalers on the same product category — agent operations infrastructure — is not a coincidence. It reflects a fundamental shift in how enterprises consume cloud services.

The numbers tell the story clearly:

• The global AI agents market is projected to reach $10.9–12 billion in 2026, with Deloitte forecasting a CAGR of roughly 53% to $45 billion by 2030.
• Agentic infrastructure now represents 17–22% of enterprise AI line items in 2026, projected to grow to 26–32% by 2027.
• Q1 2026 agent-native venture funding hit $4.7 billion (annualized to $20B+) — the largest software vertical funded since cloud-native in 2015–2017.
• 56% of enterprises now have a formal “AI agent owner” or “agentic ops” lead, up from 11% in 2024.
• 40% of enterprise applications are expected to embed task-specific AI agents by end of 2026.

The pattern is unmistakable: agent operations is following the same trajectory that container orchestration followed a decade ago. First came the runtime (agents themselves), then chaos (everyone building custom orchestration), then a scramble to own the governance and orchestration layer. The open question — and it’s a consequential one — is whether the agent world gets its Kubernetes moment (one neutral standard the ecosystem consolidates around) or fragments into per-cloud control planes that make agent governance a portability problem.

Practitioner Playbook: What to Do Right Now

Alibaba’s announcement, combined with the broader hyperscaler convergence, creates a clear set of practical implications depending on where your agent workloads currently live.

If You’re Building Custom Agents Today

The most important takeaway isn’t about Alibaba specifically — it’s that tracing, evaluation, and governance are now officially infrastructure, not application concerns. If you’re building agents with LangGraph, CrewAI, or any other framework, invest in a thin version of this layer in your own stack right now:

• Tracing — instrument every LLM call, tool invocation, and agent decision point. Use OpenTelemetry-compatible formats so you can migrate between observability providers later.
• Evaluation — build eval suites that run automatically against your agents, not just on a developer’s laptop. Measure accuracy, latency, cost-per-task, and failure modes.
• Governance — implement policy enforcement at the orchestration layer. What can each agent access? What actions require human approval? What data is off-limits?

Building this layer yourself — even in a lightweight form — means that adopting any vendor’s control plane later is a migration of convenience, not a rescue operation.

If You’re on AWS

Bedrock AgentCore is the most mature offering with real pricing and GA status. If you’re already on AWS, there’s no reason to wait for Alibaba’s offering. But keep your agent logic framework-agnostic (LangGraph or similar) rather than coupling to AgentCore-specific APIs, so you can evaluate alternatives as they mature.

If You’re Evaluating Cloud Providers for Agent Workloads

The decision framework is straightforward:

If You’re in a Regulated Industry

Nothing in Alibaba’s Agent Native Cloud announcement includes compliance certifications, data residency guarantees, or SLA commitments. For enterprises with regulatory exposure (finance, healthcare, government), there is currently nothing to evaluate from Alibaba on this front. Revisit when GA terms and legal documentation exist.

The Bigger Picture: From Cloud-Native to Agent-Native

The terminology shift from “cloud-native” to “agent-native” is not marketing — it reflects a genuine architectural evolution. In a cloud-native architecture, the fundamental unit of deployment is a container or serverless function. In an agent-native architecture, the fundamental unit is an autonomous agent that can reason, use tools, maintain memory, and collaborate with other agents.

This shift has cascading implications across the infrastructure stack:

Alibaba Cloud’s Agent Native Cloud is the most explicit articulation of this transition that any cloud provider has made. Whether Alibaba or AWS or Google ultimately wins the agent infrastructure market is less important than the fact that all of them agree the transition is happening. For practitioners, the right response is not to pick a winner, but to understand the architectural patterns and start building for them.

Frequently Asked Questions

Is Alibaba Cloud Agent Native Cloud available to use right now?

No. As of July 2026, Agent Native Cloud (specifically AgentLoop and AgentTeams) has been unveiled but has no published general availability date, no pricing, and no public preview. AgentRun, the pre-existing lifecycle management component, has been available on Alibaba Cloud, but the two new services have no disclosed timeline for launch.

How does Agent Native Cloud compare to AWS Bedrock AgentCore?

AWS Bedrock AgentCore is more mature — it’s generally available, priced (free harness, pay for compute), and has over 1 million downloads. Alibaba’s offering is more vertically integrated (covering silicon to governance in one stack) but has no pricing, no GA date, and no named customers. For teams that need to ship today, AWS is the pragmatic choice. For teams tracking where the market is heading, Alibaba’s architecture is worth studying.

Do I need to be on Alibaba Cloud to benefit from these architectural patterns?

No. The three-layer pattern Alibaba is productizing — agent runtime, agent observability, and multi-agent governance — is vendor-agnostic in concept. You can build equivalent capabilities using open-source tools (LangSmith for tracing, custom eval harnesses, policy engines) on any cloud. The value of watching Alibaba’s announcement is understanding that these are now considered infrastructure concerns, not application-level add-ons.

What is TokenWorks and why should I care?

TokenWorks unifies inference request routing, execution, compute reuse, and scheduling into a single service. If you’re running agentic systems where each task triggers multiple LLM calls, inference cost is your largest operational expense. TokenWorks addresses this by optimizing how GPU resources are shared across requests — a problem that most teams currently solve with a patchwork of vLLM, custom batching logic, and load balancers.

Will Agent Native Cloud support non-Qwen models?

Alibaba hasn’t explicitly addressed this in the WAIC announcement. Given that Alibaba’s Model Studio already hosts 200+ models and the platform is described as framework-agnostic, multi-model support is likely — but no confirmation exists as of this writing.

Key Takeaways

• Agent ops is now a cloud primitive. When four hyperscalers independently converge on the same product shape within a year, the category is no longer speculative — it’s an established layer of the enterprise stack.
• Alibaba announced two new services, not three. AgentLoop (observability) and AgentTeams (multi-agent governance) are new. AgentRun (lifecycle management) already existed and was expanded.
• Nothing is priced or dated. No pricing, no GA date, no named customer, no compliance certification. This is stated intent, not a shippable product.
• The vertical integration is unprecedented. Silicon (Zhenwu) → chip software (SAIL) → inference serving (TokenWorks) → foundation model (Qwen 3.8) → agent runtime → observability → governance. No other vendor assembled this complete a stack at a single event.
• Build your own thin agent-ops layer now. Regardless of which vendor you ultimately adopt, owning your traces, evals, and governance policies in portable formats ensures you can switch platforms later without a rescue migration.
• AWS Bedrock AgentCore is the pragmatic choice today. It’s GA, priced, and proven at scale. Alibaba’s offering is worth tracking for architectural direction, not immediate adoption.
• The Kubernetes question is unanswered. Will the agent world consolidate around an open standard (like the Linux Foundation’s agentgateway) or fragment into per-cloud control planes? Your portability decisions today depend on which outcome you’re betting on.

References

1. Alibaba Cloud Community, “Alibaba Cloud Unveils Agent-Native Innovations at WAIC 2026,” July 20, 2026.
2. Digital Applied, “Alibaba’s Agent-Native Cloud: AgentLoop and AgentTeams,” July 21, 2026.
3. CryptoBriefing, “Alibaba Cloud launches Agent Native Cloud to scale enterprise AI agents,” July 2026.
4. MarkTechPost, “Alibaba Previews Qwen3.8-Max, a 2.4 Trillion-Parameter Multimodal Model,” July 19, 2026.
5. AWS, “Introducing Amazon Bedrock AgentCore,” 2026.
6. AWS, “Amazon Bedrock AgentCore harness is now generally available,” 2026.
7. Microsoft Learn, “Transition agent security capabilities to Microsoft Agent 365,” July 2026.
8. Google Developers Blog, “Announcing the Agent2Agent Protocol (A2A),” 2026.
9. Constellation Research, “Amazon Bedrock AgentCore generally available,” 2026.
10. Salesforce Newsroom, “Salesforce and AWS Deepen Collaboration to Launch Agentforce 360 for AWS,” 2026.