Moonshot Turned Agents Into a Cloud Service: 24/7 Hosting, 140 Skills—AI Companies Now Sell Workers, Not Answers

On November 5, Moonshot AI officially launched Kimi Hosted Agents. The name sounds like yet another agent product. The approach is different: the company took the Harness—the runtime layer that handles task scheduling, tool calls, and state management—and turned it into a cloud service.

Enterprises don't build the framework or run the ops. They connect via API and tasks run continuously in a cloud sandbox, 24/7. Session history, sandbox state, and output files persist server-side. Interrupted tasks resume. Execution leaves an auditable trail.

Selling "runtime" as a separate layer is a smart cut

Agent products have been racing all year, but enterprise adoption keeps hitting the same wall: agents need to run long-horizon tasks around the clock, and building that infrastructure is genuinely hard. Session persistence, interruption recovery, sandbox isolation, execution auditing—these are infrastructure-grade engineering problems. Most companies don't want to touch them, and shouldn't.

Moonshot took the entire layer. Enterprises just define tasks. Two usage patterns, per the official release: internally, agents plug into company systems for business monitoring and ops response; externally, they embed into customer-facing products as research assistants or support agents.

One security detail deserves its own mention: Kimi Inference Hook inspects task inputs and tool returns before every model inference, intercepting private and sensitive data before it reaches the model. It's currently limited to verified enterprise users, but the direction is clear—auditing and interception capabilities are the admission ticket for agents entering the enterprise.

From selling APIs to selling agents: a business model gear shift

Shanghai Securities News called it accurately: this isn't an isolated product move. It's a key upgrade in Moonshot's B2B commercialization path—from "selling model inference APIs" to "selling agents."

The difference between the two businesses is worth calculating. With APIs, customers pay per token. However good the model, it's still a tool—pricing power stays with the customer, who can comparison-shop freely. Hosted agents sell something else: a continuously running digital worker. Framework, sandbox, skills, monitoring, audit—bundled. A tool can be swapped anytime. A worker carries switching costs: its session history, skill configurations, and business process adaptations all live on your platform.

That's the real moat logic. Platform-level competition replacing model capability competition is the road every leading lab is now taking.

Context: Kimi K3.1 Pro launched October 24—three trillion parameters, one million token context, notable gains in coding and long-horizon agent capabilities, with a Highspeed variant running 2-3x faster. Model first, hosting service second. The sequencing is deliberate.

What this means if you're doing GEO

Agent hosting changes who reads your content.

GEO so far has been about "humans asking AI": a user queries ChatGPT or Doubao, the model retrieves information and generates an answer, and brands compete to be cited. Now there's a second scenario—agents running tasks on their own. A cloud-hosted agent generating a morning briefing or tracking an IPO will proactively search the web, read pages, and verify data as the task requires. Nobody asks a question. Information gets consumed anyway.

Citation logic differs between the two environments. When humans ask, models favor fluent answers, and citations play a supporting role. When agents run tasks, retrieved information gets used: written into reports, driving decisions, triggering next actions. Wrong information in a chat misleads one person. Wrong information in an agent pipeline propagates down the task chain—and under "auditable execution" designs, the responsibility for using bad data is traceable.

This raises the bar for information quality. Brand information now has to pass three gates: machine-readable (clear structure, consistent terminology), verifiable (multi-source consistency, cross-checkable), and attributable (the brand-fact association precise enough to cite correctly).

Something concrete: turn product specs, pricing, and use cases into machine-friendly fact sheets, aligned across your website, retail listings, and industry references. Watch the retrieval habits of those eight financial agent templates—finance is where agents are landing fastest, and financial citation standards will likely formalize first.

Agents have started showing up for work. Whether your content is in their library or outside their search range—the answer is becoming the new dividing line.