Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Teams tasks

Kimi Hosted Agents: a managed harness in the cloud

Written November 8, 2026

Moonshot's Kimi Hosted Agents offers a 24/7 managed agent harness with sandboxes, audit trails, and finance templates. What it runs, and what it costs.

What it is

Moonshot AI formally launched Kimi Hosted Agents on November 5, 2026, a managed cloud service that runs AI agents for enterprises and developers without requiring them to build or maintain the underlying agent framework. The service runs on the Kimi API Open Platform and provides a 24/7 hosted environment for long-running or asynchronous work and multi-agent collaboration. A beta had already opened in September 2026 for domestic enterprise-verified users, so the November announcement formalizes a service that had been running in a limited form, and Moonshot's documentation still shows a Beta label on the service.

The pitch is the unglamorous part of agents. A production agent needs sessions, tool execution, sandboxes, state, retries, scheduling, and failure recovery. That operational layer is what most agent projects build themselves over months. Hosted Agents packages that layer as a hosted service: teams configure agents that run on Kimi's infrastructure, and the platform stores complete session history as events, runs built-in tools in isolated cloud sandboxes, handles execution retries and scheduling, and resumes automatically after a platform failure. The service is stateful. Session history, sandbox state, and output files persist on the server, and its event stream supports resumption with a Last-Event-ID header after a disconnected client reconnects. Each thread keeps separate status, history, and event streams, so operators can inspect and audit individual subtasks.

Controls and templates

The control layer is isolated sandboxes, network-access controls, configurable operating boundaries, sessions an operator can stop at any time, and execution records that can be exported for audit. An optional component, the Kimi Inference Hook, lets a company's own safety service, or one supplied by a third party, inspect task inputs and tool returns before each model call, so only content that passes the enterprise's own rules reaches inference. The hook is offered selectively, to enterprise-verified users only.

The catalogue is built for office work rather than demos. The platform lists more than 140 skills and more than 30 plugins covering web search, professional data queries, office documents, and content generation, and teams can attach their own tools, data sources, and MCP services. The documentation lists 12 official agent templates: eight finance templates covering modelling, deep research, morning briefs, IPO tracking, and portfolio monitoring, plus game operations, a player advisor, SRE incident response, and Discord community operations. Each template can be copied into a project space and wired to private databases. Moonshot positions the service for internal business monitoring and operations response, and for customer-facing investment research and customer-service assistants.

There is also a self-hosted option, in which sandbox tool execution moves to a customer-managed host while Kimi's platform continues to handle model calls, orchestration, and session management. Moonshot is explicit that this is not equivalent to private deployment, because tool inputs and execution results still pass through the platform's model-inference interface, and the self-hosted version does not support IPython, image-search tools, or save_artifact.

What it costs

Pricing is pay-as-you-go in three parts, per Kimi's pricing page: sandbox runtime at ¥0.072 per hour (about one US cent), model inference charged at the selected model's token rates, and plugin charges based on call counts and each plugin's unit price. An agent becomes a metered workload (sandbox hours plus inference plus plugin calls) rather than a seat licence. Which model the service runs, and which regions can use it, were not stated in the launch coverage read for this page; the earlier beta was limited to domestic enterprise-verified users, and the launch coverage does not say whether that scope has widened.

What to do first

  1. Start from a template, not a blank agent. Copy one of the 12 official templates into a project space and connect it to a real internal data source. A template that works against your own data answers the only question that matters, whether the service does useful work, faster than a custom build.
  2. Price one real workload before committing. Run a representative task, then add up the three metered parts: sandbox hours, model tokens, plugin calls. The sandbox rate looks small; the total is what a budget needs.
  3. Decide who needs the Inference Hook before you need it. It is gated to enterprise-verified users. If your agent touches regulated content, confirm eligibility first. It is not a switch you flip later.
  4. Do not read self-hosted as private. Tool inputs and results still pass through Kimi's model-inference interface. If the requirement is that data never leaves your machines, this is not that product.

Limits

Related playbooks

Sources

Back to: Small-team playbooks hub · Also: Kimi K3.1 Pro weights are out: what 1.56 TB buys · OpenAI's Agents API: the harness is now hosted · Baseten Grid routes coding-agent traffic