Cloud vendors now sell the agent platform layer

After 22 September 2026 the buying question for enterprise agents changed. Cloud vendors turned harness, context and governance from integration projects into products — and in the same keynote cycle built distribution for rival models and rival harnesses. Feature lists no longer separate platforms. What remains to ask is where the execution boundary lives and whether the evidence travels with you.

Apsara turned the agent platform layer into products

At the Apsara 2026 technical keynote, Alibaba Cloud CTO Li Feifei split Agentic Cloud into three build scenarios — Model, Harness and Context — and launched AgentCore, an enterprise agent build-and-governance platform: long-running tasks, retries, checkpoint recovery and async execution, plus a secure execution environment, resource isolation, identity and permissions, and full-chain audit, with models, MCP Servers and Skills managed as one inventory (TechWeb on-site report, 22 September 2026).

Around the same layer the vendor shipped a set of execution products: Agent Sandbox for elastic, E2B- and K8s-compatible execution environments; Agentic OS as an agent-native operating system; Agentic Computer as a long-lived cloud workstation with Computer Use and GUI operation; and Context Engine, which converts enterprise-wide data into live agent context. The vendor-stated figures include 99% task completion, 70% lower TCO, 100,000 sandbox creations per minute, and a 50-percentage-point accuracy gain for Agent Context.

Every one of those numbers comes from Alibaba Cloud’s own conference with no third-party reproduction. One of them is internally inconsistent: the international product documentation for Agent Sandbox states up to 15K sandboxes per minute, roughly 6.7x below the keynote figure, and neither source states its measurement scope. So this article cites no single throughput number and records only the structural fact: sandboxing, identity, permissions and audit are now written into the vendor’s product definition.

The more interesting half: an entry point for rival harnesses

At the MaaS & Agent forum that afternoon, Alibaba Group strategy vice president Wen Zheng launched Token Plan and Agent Studio. Token Plan covers Qwen, Kimi, GLM and DeepSeek under one subscription and natively connects mainstream coding tools and agent frameworks including Qoder, Cursor, Codex and OpenClaw (DOIT, 22 September 2026).

The commercial terms of that product definition are undisclosed, so it cannot be read as the cloud vendor reselling rival models. The product shape alone settles one thing: a replaceable harness is now the vendor market’s default assumption. A cloud vendor building an official entry point for someone else’s harness is telling you it does not expect customers to be locked to one.

For buyers this compresses two of the differentiators they used to lean on. Governance capability becomes table stakes, and the harness becomes a swappable part. Two positions remain defensible: runtime evidence that is independent of the vendor, and a private deployment path.

A verifiable side note: data residency is already a selling point

Alibaba Cloud’s technical solution catalogue now lists the Hermes Agent solution and names enterprises that want to run agents in a private environment with data staying in their own domain as a target customer, with an ECS one-click deployment plus Bailian model service as the path (curl returned HTTP 200 on 22 September 2026). Our own runtime sits on that path, so this is an internally verifiable fact rather than a relayed claim.

It shows two things. First, private deployment and data residency are now demands the vendor must meet proactively, not clauses a customer has to negotiate. Second, neutrality cannot be proved by vendor endorsement: which vendor’s catalogue lists a runtime, and whether that runtime is independent of that vendor, are separate questions. The evidence a buyer needs comes from run records, not from an ecosystem page.

Our judgment: four questions replace the feature list

Feature lists have converged. Buying criteria should move to questions that can be verified:

1. Where does the execution boundary live. Sandboxing and permissions enforced in the architecture, or an audit log written after the fact. The difference only shows up after an incident, but it can only be asked before purchase.
2. Can the evidence be exported. Audit records in a vendor-independent format, with models and harnesses landing on the same record. Evidence you can only read inside one console dies when you switch platforms.
3. Does data leave your domain. Whether a private or local deployment path actually exists and has customers running on it, rather than a checkbox on a solution page.
4. What does switching cost. If the vendor assumes the harness is replaceable, put replacement cost in the contract: migration tooling, reusable configuration, retained history.

One question for buyers

If your agent platform comes from a cloud vendor, answer this: what stops you from moving today — technical dependency or evidence dependency? If it is the second, what you need is not another feature comparison but a run record you can take with you.

References: TechWeb, “Alibaba Cloud CTO Li Feifei: building Agentic Cloud around Model, Harness and Context” (22 September 2026, 13:07; AgentCore governance scope, Agent Sandbox, Agentic OS, Agentic Computer, Context Engine and the 99% / 70% figures are all vendor-stated), www.163.com/dy/article/L7ECIPB405119MAU.html (curl HTTP 200) | DOIT, “Qwen AI platform upgrades to full-stack services” (22 September 2026, 15:06; Token Plan model and tool coverage, Agent Studio, 50+ atomic APIs, One Key MCP, 1200s to 70s model start, 38% lower first-token latency, up to 95% Prompt Caching savings and 99.9% SLA are vendor-stated), www.doit.com.cn/ai/851657236795461.html (curl HTTP 200) | MyDrivers / Kuaikeji, “Qwen Office releases Enterprise Context” (22 September 2026, 15:16; enterprise context, digital-employee identity fields, and the pre-execution, in-flight and post-incident controls of the security centre, including Chen Yusen’s remarks), m.mydrivers.com/newsview/1153075.html (curl HTTP 200) | Alibaba Cloud technical solutions, “Hermes Agent” (target customers include enterprises requiring data residency; ECS one-click deployment plus Bailian), www.aliyun.com/solution/tech-solution/hermes-agent (curl HTTP 200) | Alibaba Cloud ECS documentation, same solution, help.aliyun.com/zh/ecs/user-guide/hermes-agent (curl HTTP 200) | Agent Sandbox international product documentation (source of the throughput discrepancy: up to 15K sandboxes per minute), www.alibabacloud.com/help/tc/cs/user-guide/agent-sandbox (curl HTTP 200). Vendor-stated figures are not third-party reproduced; recorded here as vendor claims only, not as procurement commitments.