AgentOS from openJiuwen allows agents to independently create, install, and roll back plugins

Moonshot AI has launched Kimi Hosted Agents—a cloud service that packages the agent runtime environment, or 'harness,' as a managed product. This allows enterprises to avoid the need to build and maintain this infrastructure layer themselves, according to information from IT Home and the company's documentation on its open platform.

This offering is positioned as 'Harness as a Service.' Teams can configure agents on the Kimi infrastructure and receive a 24/7 environment for executing long-running tasks and collaborative work among multiple agents.

Enterprises can direct agents to internal systems for monitoring, research, and operational activities, or integrate them into customer-facing products as support or research assistants. Kimi Hosted Agents provides isolated sandboxes with network access controls, session boundaries that operators can stop at any time, and execution logs that can be exported for auditing. According to IT Home, the sandbox costs 0.072 yuan per hour plus model and plugin fees.

The platform currently includes over 140 skills and more than 30 plugins covering web search, professional data queries, office documents, and content generation. Developers can also connect their own tools, data sources, and MCP services. An additional security component, Kimi Inference Hook, allows the enterprise's proprietary or third-party security service to check task input data and tool output results before every model call, ensuring that only company-compliant content enters the inference process. This hook is currently available only to verified enterprise clients.

Official documentation also describes options for self-hosting environments, where sandbox tools run on customer-controlled machines while orchestration remains on the platform. In addition to the managed runtime, Moonshot is releasing the first series of official agent templates. IT Home notes eight financial templates covering modeling, deep research, morning briefings, IPO tracking, and portfolio monitoring; the documentation mentions a broader initial set of 12 templates, including game management and SRE incident response.

This launch expands Kimi's product line from models and desktop agents to a managed runtime, which Moonshot states the company will support and update. This occurred one day after Huawei Ascend announced support for zero-day inference for the open-weight Kimi K3.1 Pro model, highlighting a broader effort to simplify the deployment of Kimi agents in production environments.

openJiuwen, an open agent community supported by Huawei, has released Hyper-Agent—a swarm architecture that moves agents from isolated chat sessions into real group work chats, transforming them into permanent digital employees. This was reported by IT Home as part of the community's publication.

The functionality is available in openJiuwen WorkSwarm, whose code is open-sourced on GitHub under the name JiuwenSwarm. The main change lies in altering the principle of agent connection to sessions. Instead of an agent being tied to a single conversation, the Persistent Agent retains its identity and state, continuously performing tasks while sessions become a base resource that the system distributes, schedules in parallel, and resumes while preserving context.

The system includes a gateway connecting instant messaging applications and a web interface, along with an AgentServer responsible for planning agents and work sessions. According to the community, this development builds upon previous work on Persistent Sessions and allows one agent to perform multiple tasks across various third-party group chats.

In WorkSwarm, users create these agents through an employee center, using either ready-made expert templates or creating them from scratch. When configuring, responsibilities, working style, delivery requirements, skills, connectors, knowledge sources, and even a dedicated email address are specified. When the context window approaches its limit, the agent switches to a new one, passing a summary of goals, progress, and next steps, and retrieves archived history if necessary.

There are two operating modes: passive mode only responds when mentioned, and adaptive mode decides whether to ignore a message, reply, or intervene based on its role and the flow of the conversation. An activity panel displays each agent's activity, token usage, and recent response time.

The same agent can operate in office applications such as Feishu and DingTalk, accepting a request in one channel and providing the result in another. Approval is required for information exchange between groups: users verify the source group, the recipient, and the content, after which they grant permission once or permanently, which can be revoked later.

Team members can also run WorkSwarm on their own devices, allowing agents on cloud servers and personal PCs to collaborate in the same group using locally authorized files and tools.

For larger projects, WorkSwarm utilizes the SwarmEngine group chat, which unites users, persistent agents, individual experts, and expert teams in one space. The expert team has a leader who takes tasks, reports progress, and organizes deliverables, while team members work in the background. Experts can be drawn from third-party agents like Codex and Claude Code and mixed with internal digital employees.

The community notes that future plans include the possibility of coordinating personal and cluster computing resources, active learning, and failure recovery to ensure longer-term collaboration.

NetEase Yidun, the content safety and risk management division for NetEase's enterprise services, unveiled its native AI product suite at the World Internet Conference Light of Internet Expo in Wuzhen on November 3, 2026, according to Sina Tech and IT Home.

The company asserts that security must shift from controlling what AI says to controlling what it does, as agents gain access to corporate systems, tools, and knowledge bases.

Zhu Haoci, Deputy General Manager of NetEase Enterprise Services and head of Yidun, presented a dual model: using AI to enhance security operations and applying security tools to constrain AI functionality within defined boundaries. He noted that current risks include AI content generation, deepfakes, automated fraud, model jailbreaks, confidential data leaks, and agents performing actions outside their permissions.

The product suite consists of three levels. The first level covers AI content and data security, including AI content analysis and moderation, as well as the CMA moderation agent. The second level focuses on interaction and identity security, encompassing business risk control, device fingerprints, behavioral CAPTCHA, and DeepGuard anti-spoofing protection. The third level is responsible for model and agent security, providing mechanisms for large model protection, an agent security platform, and agent supply chain verification.

The system received first-version certification at the exhibition. At the agent level, protection mechanisms filter model inputs and outputs, while Agent Guard verifies agent actions against user intentions, ensures permission compliance, flags risky tool calls, and blocks behavior.

The agent security platform detects utilized AI assets, scans Skill and MCP supply chains, and protects against confidential data leakage. It targets shadow AI—tools used by employees without IT department awareness—by offering blocking, auditing, alerting, and data masking options.

Yidun provided examples of other product effectiveness. CMA integrates an internal content safety model with specialized smaller models and operates 24/7. The company reported that when analyzing complaints for one client processing over 2.5 million tickets monthly, CMA's initial review combined with human checks reduced the average review time from 35.7 minutes to 12.3 minutes with 99.2% accuracy. DeepGuard, which searches for improbable content rather than just known forgery traces, received top marks in the Chinese Academy of Information and Communications Technology's anti-face spoofing assessment and demonstrated 99% completeness and 96% accuracy against screen replay attacks in phased testing.

At a separate forum, Rao Xiaoyan, Head of Yidun's Content Security Department, stated that malicious instructions can be hidden in web pages, files, APIs, tools, or MCP responses, so systems must differentiate between task data and behavioral commands. She described implementing injection defenses both before model calls and before context assembly, as well as output checking during streaming: questionable content is flagged for review, and high-risk content is rejected.