Share with your CTO
Moonshot AI is making a direct play for enterprise infrastructure budgets with Kimi Hosted Agents, a cloud platform that lets businesses deploy and manage AI agents without building the underlying runtime themselves. The service bundles over 140 skills, 30-plus plugins, sandboxed execution environments, and an optional pre-inference security hook called Kimi Inference Hook. Pricing separates infrastructure (0.072 yuan per sandbox hour) from model calls and plugin usage. A self-hosted sandbox option targets regulated industries. This is Moonshot’s clearest move yet from consumer chatbot to enterprise infrastructure vendor.
What this means for your business
The story lands differently depending on whether you’re already running agentic workloads or still in proof-of-concept. If your team has spent the last six months building custom agent harnesses, the runtime scaffolding that manages what tools an AI agent can call, what tasks it queues, and how it connects to external services, Moonshot is now selling that layer as a commodity. That’s either a threat to your build-it-yourself roadmap or a reason to stop building it yourself.
The “Harness as a Service” framing is the structurally interesting move here. Every serious cloud vendor is racing to own the agent execution layer because whoever hosts the harness sees the task graph, the tool calls, and the data flows. Moonshot is packaging that layer with security controls, audit logs, and MCP (Model Context Protocol, an emerging standard for letting AI agents connect to external data sources and tools) compatibility. It’s a credible stack for mid-market enterprises that lack the platform engineering headcount to wire this together themselves. The risk is vendor lock-in at a layer most CTOs haven’t thought carefully about yet, because the harness is invisible until something goes wrong.
The Kimi Inference Hook is the detail worth watching most carefully. Letting an external security system inspect agent inputs and tool outputs before they reach the model is a genuinely differentiated control, and the fact that it’s gated to enterprise-verified customers signals Moonshot knows its highest-value buyers are the ones with existing security review requirements. If that feature matures and gets third-party certification, it changes the conversation with procurement and legal teams who’ve been blocking agent deployments on data-exposure grounds. I’d revise that optimism if the feature stays China-infrastructure-only and never reaches global sovereignty requirements.
Concept deep-dive: Agent Harness
An agent harness is the runtime layer sitting between an AI model and the outside world, think of it as the operating system for an AI agent. It decides which tools the agent can call, sequences multi-step tasks, manages memory across a session, and enforces boundaries on network and data access. Without a harness, deploying an agent means writing that orchestration logic yourself. Hosting it as a managed service shifts that engineering burden to the vendor, which is exactly the trade-off Moonshot is now selling.
Based on reporting from Moonshot AI Launches Hosted Agents for Enterprise AI Operations Hubs, originally published 2026-11-04 23:35:00.
