Kimi Turns Long-Running AI Agents Into a Managed Enterprise Service

Moonshot AI’s Kimi formally launched Kimi Hosted Agents on November 5, 2026, turning its agent harness into a managed cloud service for enterprise AI work. The platform targets long-running tasks, asynchronous work, and collaboration between multiple agents, giving developers a hosted way to run systems that need more than a single model response.

The launch follows a Hosted Agents beta listed in September 2026 for domestic enterprise-verified users. In other words, the November announcement formalizes a service that had already appeared in beta and now comes with a clearer set of tools for building, monitoring, and controlling agent workflows.

A hosted runtime for work that keeps moving

Kimi describes the agent harness as the operational layer behind an AI agent. It manages sessions, tool execution, sandboxes, and failure recovery, rather than leaving each organization to build those pieces from scratch. Hosted Agents packages that layer as a prebuilt, configurable service running on Kimi’s infrastructure.

Kimi positions the service for internal business monitoring and operations response, along with customer-facing investment-research and customer-service assistants. These uses share one challenge: the work may continue after the first request, require several steps, or need a response after a delay.

The hosted platform stores complete session history as events and keeps session history, sandbox state, and output files on the server. It also runs built-in tools inside isolated cloud sandboxes, handles execution retries and scheduling, and automatically takes over after a platform failure. Kimi says execution resumes in a state consistent with the point before the interruption.

Custom tools still run on the customer’s client. Developers can work through a web console, REST API, or the kha command-line interface. Console access requires an account login, while CLI and cURL workflows use an API key.

The event stream also supports reconnecting after a client loses its connection. A client can resume with a Last-Event-ID header, while delivered files can be listed and downloaded as artifacts. That combination gives teams a record of what happened, along with a way to recover work without starting a session over.

Multi-agent collaboration with limits and oversight

Hosted Agents supports multi-agent work through a coordinator agent. The coordinator delegates subtasks to child threads within the same session, and a delegation configuration can contain between one and 20 agent entries.

Teams can choose between synchronous and asynchronous delegation. Synchronous delegation waits for a child thread to report, become idle, or terminate. Asynchronous delegation returns as soon as the child starts, with the report delivered later.

The structure has clear boundaries. Child threads cannot create another level of child threads, and they cannot communicate directly with one another. Each thread keeps its own status, history, and event stream, which supports inspection and auditing of individual subtasks.

Kimi’s control layer includes independent sandboxes, network-access controls, configurable operating boundaries, session interruption, execution records, and audit export. Kimi Inference Hook can connect to enterprise-built or purchased security services and inspect task inputs and tool returns before each model inference.

The enterprise defines the rules, and the hook allows or rejects the content. Kimi is offering that hook selectively to enterprise-verified users, matching the service’s focus on controlled deployment rather than an open, unmanaged runtime.

Hosted convenience, self-hosted execution

Organizations that want a custom harness can still use Kimi’s direct model-inference API. That option gives them finer control, but it also means they must build the operational layer themselves instead of using the hosted service.

Kimi also offers a self-hosted environment to enterprise-verified users. In that setup, sandbox tool execution moves to a customer-managed host, while Kimi’s platform continues to handle model calls, agent orchestration, and session management.

The split gives enterprises two ways to approach deployment. Hosted Agents keeps the runtime and its operational controls on Kimi’s infrastructure, while self-hosting places sandbox execution on the customer’s host without moving every part of the platform.

Why context and memory remain central

Kimi’s launch arrives alongside Redis’s next chapter of Redis Iris, a real-time context engine for AI agents. Redis Iris now offers six customizable capabilities that can work together or independently: Redis Intent Router, Redis Agent Playbooks, Redis Agent Memory, Redis Context Gateway, Redis Lens, and Redis Agent Mesh.

Redis describes Iris as giving agents the memory, context, and messaging they need to work reliably in real-world settings. Intent Router, Agent Playbooks, Agent Memory, and Context Gateway are generally available, while Lens and Agent Mesh are in beta. Redis Iris first launched in May 2026, and customers have been putting it to work in production since then.

“Agents need the same foundations as apps and microservices: live app data, runtime state and service communication,” Rowan Trollope, CEO of Redis, said.

June Lee, CTO of Paradigm, described the company’s experience this way: “Everything you see in Paradigm is stored in Redis. It is the memory and orchestration layer that lets our agents work in real time, at a scale that wasn’t possible before,”

Redis will host a live first look at the release with its engineering team on Nov. 12 for the Americas and Nov. 17 for EMEA. Redis Iris is available today.

Elsewhere in the AI landscape, Argon Fieldnotes launched an independent Gemini 4 guide to help developers and technical decision-makers assess Google’s Gemini 4 Argon announcement. The guide organizes available evidence around practical adoption questions and is described as an editorial resource, not a Google product, an API provider, or a way to obtain early model access.

Google announced Gemini 4 Argon on September 30, 2026. Initial access is limited to trusted cyber defenders through the Fairwind Program, with a wider rollout planned.

Stay connected via Google News

Follow us for the latest AI News