NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Moonshot Now Rents the Unglamorous Half of an AI Agent, Billed by the Hour

Kimi Hosted Agents sells the harness around the model — sessions, sandboxes, retries — for RMB 0.072 an hour. Selling uptime, not tokens, changes the economics.

2026-11-16 · 1180 words · NeuroAI
Moonshot Now Rents the Unglamorous Half of an AI Agent, Billed by the Hour

Ask anyone who has moved an AI agent (智能体) from a demo to production what broke first, and you rarely hear "the model". You hear about the session that died at hour six and could not be resumed. The sandbox that leaked a file. The retry loop that fired twice and sent the same email to two thousand customers.

Moonshot AI is now selling a fix for all of that, priced by the hour.

On November 5, 2026, the company behind Kimi formally launched Kimi Hosted Agents, a managed service that runs the whole harness (运行框架) around a model — session state, cloud sandboxes, retries, scheduling, multi-agent delegation — and hands it to enterprises that would rather not build it themselves. The service had been in beta since September for verified domestic enterprise accounts.

What is actually being sold

Strip the marketing and the product is a stateful execution environment.

Session history is stored as events, so a client that disconnects can reattach and keep going. Sandbox state and output files persist server-side. Tools execute in isolated cloud sandboxes with network controls. Retries, scheduling and failure recovery are the platform's problem — Moonshot says it resumes in a state consistent with pre-interruption behaviour — and execution records export for audit.

Developers get in three ways: a web console, a REST API, and a command-line tool called kha that runs inside Kimi Code, Claude Code and Codex.

Around that core sits a catalogue meant to shorten the distance from trial to deployment: more than 140 skills, more than 30 plugins covering web retrieval, professional data queries, office documents and content generation, plus MCP services and custom tools for internal systems. Twelve agent templates ship with it, and eight of those twelve are financial, including modelling, deep research, morning reports, IPO tracking and portfolio reviews.

That template mix is the tell. Moonshot is not pitching a general assistant. It is pitching the work that happens at 7am before a trading desk opens.

The price is not the token

The large model (大模型) business has had one meter for three years: tokens. Hosted Agents adds two more.

The bill is the sum of three parts — sandbox runtime at RMB 0.072 per hour (≈US$0.01 / HK$0.08), model inference at the chosen model's token rates, and plugin calls at unit prices that run from RMB 0.001 to RMB 0.45 each (≈US$0.00014 to US$0.063 / HK$0.001 to HK$0.49). Web search is RMB 0.03 a call; image search RMB 0.045; some financial data plugins RMB 0.35 or RMB 0.45.

Do the arithmetic and the architecture of the business becomes visible. A sandbox running continuously for 30 days — 720 hours — costs about RMB 51.84 (≈US$7.2 / HK$56) in runtime. That is almost nothing. The money is in the tokens the agent burns while it sits in that sandbox and in the plugin calls it makes, especially the expensive financial feeds.

This is a different economic shape from a chat API. Chat bills for what a user asked. Hosted agents bill for how long software was awake and what it touched. Moonshot's usage reports break consumption down by agent project, which tells you what buyers worry about: not price per token, but which project is quietly overrunning its budget.

A coordinator, not a swarm

The multi-agent design is the most opinionated part of the launch, and it runs against fashion.

Hosted Agents uses one coordinator agent that delegates subtasks to child threads inside the same session. A delegation configuration can hold between one and 20 agent entries. Delegation can be synchronous — the parent waits for a child to report, go idle or terminate — or asynchronous, returning as soon as the child starts. Each thread keeps its own status, history and event stream.

Then the constraints. Child threads cannot spawn their own children, and they cannot talk to each other.

Read that as a design position rather than a limitation. A two-level tree with no lateral chatter is dramatically easier to audit, bound and debug than a mesh of peers negotiating among themselves. Every subtask becomes individually inspectable. Every failure has one obvious parent to blame. If you have ever tried to reconstruct why a swarm of eight agents produced a wrong number, the appeal is immediate.

The compliance door

Buried deeper in the documentation is the feature security teams will actually read twice: Kimi Inference Hook. It lets an enterprise attach its own safety service — built in-house or bought — to inspect task inputs and tool returns before each model inference, then allow or reject the content according to rules the enterprise writes.

That is pre-execution interception, not after-the-fact log review, and it is the difference between an agent that can touch a core business system and one that cannot. The hook is gated to enterprise-verified users.

Moonshot also offers a self-hosted option that moves sandbox tool execution onto a customer-managed host while orchestration, model calls and session management stay on its platform. The company is blunt about what that is not: because tool inputs and results still pass through its model interface, it is not private deployment. The variant also drops IPython, image search and artifact saving.

Why this is a harder business than it looks

Selling uptime is a different trade from selling intelligence. Moonshot is strong on the model-to-harness fit — few rivals own both the model and the harness, as far as I can tell — but it owns no cloud infrastructure. Its rivals are the clouds that already host enterprise workloads and can bundle compute into the same invoice.

The launch is also China-first. At the time of writing, the international Kimi platform carried no Hosted Agents pages, and the documentation checked for this article was in Chinese. Enterprises outside China can read the pitch but cannot easily buy it yet.

Honest limitations

  • Capability and pricing figures here are Moonshot's own. No independent benchmark of the hosted service was found, and none is likely soon — managed runtimes are hard to benchmark fairly.
  • The runtime fee is nearly trivial only if the agent is idle. A long-running agent burning tokens around the clock is a token bill with a tiny hosting surcharge attached, not a cheap product.
  • The no-lateral-communication rule will frustrate teams whose workloads genuinely need peer negotiation. Moonshot has chosen auditability over flexibility, and that trade-off is real.
  • This article did not test the service. Pricing is per published rates and may change.

What readers can do now

  1. Price your agent in three lines, not one. Sandbox hours, tokens, plugin calls. If a vendor quotes you a per-token number for an agent that runs all day, you do not have a quote yet.
  2. Audit the topology before the model. Ask any agent platform how many levels of delegation it permits and whether children can talk to each other. It predicts your debugging experience more reliably than any benchmark score.
  3. Treat the inference hook as the buying criterion in regulated industries. Pre-execution interception with customer-defined rules is a procurement gate, not a feature bullet.

Related coverage

Cite this article

NeuroAI (脑机网), "Moonshot Now Rents the Unglamorous Half of an AI Agent, Billed by the Hour," 16 November 2026, https://neuroai.site/a/na-model-kimi-hosted-agents-runtime

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories