Moonshot AI announced on November 5 that Kimi Hosted Agents is now generally available, turning the harness that keeps AI agents running into a cloud service. Enterprises no longer need to build and maintain their own runtime framework: the platform provides a managed environment that operates around the clock, supports long-horizon task execution and lets multiple agents collaborate. Moonshot opened a beta with session management, tool execution, sandboxing and failure recovery back in September.
The harness is the unglamorous layer that determines whether an agent actually survives contact with production: context management, tool calls, permission control, interruption recovery. Moonshot's pitch is that building this infrastructure reliably can take a company months of engineering, so it now sells the layer itself instead of only the model behind it.
On the tooling side, Hosted Agents ships with more than 140 skills and over 30 plugins covering web search, professional data queries, office document processing and content generation. Developers can also wire in their own data sources and business systems, and configure agents through the console, API or command-line tools.
Security controls get their own section: an isolated sandbox per workload, network access permissions, session termination and execution audit logs, plus usage reports for cost management. A separate component called Kimi Inference Hook, currently limited to enterprise-certified accounts, lets companies run their own checks on task inputs and tool outputs before anything enters the model's inference flow, with the customer defining which content gets through.
Moonshot is leading with vertical templates: eight finance-focused agents covering investment research modeling, deep research, daily morning notes, IPO tracking and position monitoring. Companies can copy a template into their own project space, attach internal research reports and private databases, and run it as a persistent business agent inside the hosted environment.
The release is also a statement about where Moonshot's revenue should come from: agents rather than tokens. The company's current flagship, Kimi K3.1 Pro, launched October 24 as a 3-trillion-parameter MoE model with roughly 1 million tokens of context, and its weights have now landed on Hugging Face. Citing Arena's published agent leaderboard based on more than 17,500 real user sessions, Moonshot says K3.1 Pro ranks first among open-weight models and 13th overall, with a net improvement rate of +5.1 percent, Bash reliability of +7.9 percent and steerability of +6.5 percent; those rankings come from the third-party platform as relayed by the company.
Moonshot is entering a lane that is getting crowded fast. AWS sells Bedrock Managed Agents, and ByteDance's Doubao Work recently opened its enterprise agent to third-party models, including Kimi's own. The bet is that most enterprises would rather rent a maintained runtime than assemble one, and that whoever owns the harness layer owns the relationship with the agent, not just the inference bill.
Comments (0)
Log in to join the discussion
Log InNo comments yet