火山引擎Agent Infra升级，让Agent真正进入企业工作流

Volcano Engine's FORCE Conference: Making Agents Actually Work for Enterprises

Enterprises want to use Agents, but not for "chatting"—they want Agents that can actually get work done, integrate with existing systems, and deliver measurable ROI.

Volcano Engine (ByteDance's cloud arm) unveiled a suite of "Agent Ready infrastructure" at its recent FORCE conference, targeting exactly that. The solution has three layers: bottom layer is AI cloud (providing compute), middle layer is Agent infrastructure (providing capability modules), and top layer is enterprise AI applications (ready-to-use Agent products).

The core logic is pretty clear: for Agents to truly enter enterprise workflows, you need more than just a powerful LLM—you need a set of infrastructure that makes Agents "reliable, controllable, and measurable."

AgentKit Upgrade: Giving Agents an "ID Card" and a "Sandbox"

AgentKit is the core module of Volcano Engine's Agent infrastructure. This upgrade focuses on four components: Identity, Runtime, Sandbox, and Evaluation.

**Identity** answers "who is this Agent?" In enterprise environments, Agents don't run anonymously—they need their own "ID card" to access corresponding systems and data. AgentKit's Identity module has already integrated with thousands of enterprises' identity systems (including Feishu, DingTalk, WeCom, and various SSO solutions). Enterprises don't need to rebuild identity management from scratch—just plug in their existing system.

**Runtime** answers "how does the Agent run?" Enterprise Agent tasks often can't be completed in seconds—they may need to run for minutes, hours, or even longer. AgentKit's Runtime supports long-horizon tasks (tasks that run for extended periods) and supports up to 120,000 sandboxes running Agent tasks concurrently at minute-level scaling. That's a pretty wild number—meaning an enterprise can have 120,000 sandbox environments running Agent tasks simultaneously, with minute-level scaling.

**Sandbox** answers "will the Agent break things while running?" A sandbox isolates the Agent, letting it run in a controlled environment—even if something goes wrong, it won't affect the main system.

**Evaluation** answers "how is the Agent doing?" Without evaluation, there's no optimization. AgentKit provides an evaluation framework for enterprises to measure the effectiveness of their deployed Agents.

ArkClaw Enterprise: Turning Agents Into "Employees"

Infrastructure alone isn't enough—enterprises also need ready-to-use Agent products. That's what ArkClaw Enterprise does.

ArkClaw Enterprise integrates several components: Agent Square (various pre-built Agents enterprises can pick from), Skill Center (capabilities Agents can call), and Enterprise Knowledge Base (letting Agents access internal corporate documents and data).

For access control, ArkClaw supports IDP/SSO/OAuth and integrates with IM tools like Feishu and DingTalk—meaning employees can talk to Agents directly within Feishu, without opening a separate AI tool window.

Real-World Cases: Haidilao and Skyworth

Enough theory—what can this stuff actually do? Volcano Engine shared two compelling case studies.

**Haidilao** (the hot pot chain): store operations Agents compressed work that used to take hours of manual processing down to minutes. Specifically, manual follow-up time decreased by 70%, and inspection satisfaction increased by 50%. Haidilao's store operations involve大量 inspection, reporting, and task assignment—tasks that are ideal for Agents, because the rules are clear and they're highly repetitive.

**Skyworth**: using ArkClaw Terminal Edition to build an AIOS, Token consumption decreased by 50%, supporting millions of terminals. Skyworth's smart TV terminal count is massive—enabling every terminal to use AI capabilities while controlling costs requires highly efficient and scalable Agent infrastructure.

Where's the Real Value?

I think the smart part of Volcano Engine's approach is: they're not selling "a really powerful AI model"—they're selling "a complete solution that lets you actually use Agents."

Enterprise customers don't care how many parameters your model has or which leaderboards it tops—they care about: can I integrate Agents into my business systems? Is it secure after integration? How high is the running cost? Who's responsible if something goes wrong?

Volcano Engine's solution targets these very practical questions. AgentKit addresses reliability and controllability; ArkClaw addresses usability and integration; the underlying AI cloud addresses compute power and cost.

For large and medium enterprises considering deploying Agents internally, this solution is worth serious study. Of course, ultimate cost-effectiveness depends on specific business scenarios and scale.