Introduction
On November 5, 2026, Moonshot AI officially launched Kimi Hosted Agents under its "Harness as a Service" model, marking a new phase of competition in China’s enterprise agent API market. The release adds another major contender to a field already populated by OpenAI, Anthropic and Cursor, and accelerates the industry’s structural shift from raw model capability competition to full-stack agent runtime and orchestration capability.
For engineering and product teams, the rise of hosted agents represents a fundamental change in how AI systems are built and operated. Where earlier-generation model APIs delivered only single-turn inference, modern agent platforms provide a complete runtime substrate — commonly called a harness — that handles session state, tool execution, sandbox isolation, failure recovery and persistence out of the box. This transition dramatically lowers the barrier to deploying production-grade autonomous workflows, but it also introduces a new set of architectural decisions, most notably whether to build and operate a self-hosted harness or adopt a fully managed service.
1. The Agents API Emerges as the Core Battlefield
The first wave of large language model API competition revolved around inference speed, context window size and per-token pricing. As agentic systems move from experimental prototypes to core production workloads, the competitive frontier has shifted up the technology stack. A modern agents API does not merely expose a chat completion endpoint; it delivers an end-to-end runtime environment for autonomous multi-step workflows.
Four major platforms now define the global and domestic competitive landscape, each with distinct positioning and technical tradeoffs.
OpenAI Agents API
OpenAI was among the first vendors to formalize this product category with the Agents API, built natively on its Codex harness. The platform exposes Session, Environment and Events primitives, and gives customers the option to run sandbox environments on OpenAI’s managed infrastructure or self-host them within their own networks. Its broad tool integration ecosystem and native Codex code execution capabilities make it the most feature-complete global offering.
Anthropic Claude Agent Ecosystem
Anthropic has advanced its agent strategy through the Claude Agent SDK, alongside managed agent products integrated into the broader Claude ecosystem. The platform emphasizes native reasoning reliability and long-context stability as its core differentiators, targeting complex, multi-step enterprise workflows where accuracy and consistency take priority over maximum feature breadth.
Cursor Managed Agents
Cursor entered the space with its own Managed Agents product and an open SDK that supports custom file-based agent definitions. Engineered primarily for software development use cases, it offers deep integration into the Cursor editor ecosystem and compatibility with both self-hosted and third-party harness runtimes. Its positioning leans toward engineering teams that value high customizability and native developer tooling.
Kimi Hosted Agents
Kimi’s entry brings a China-market focused offering with deep optimization for Chinese language scenarios and pre-built vertical industry templates, with finance as its flagship vertical. The platform ships with ready-to-use financial agent templates covering research, compliance, fund analysis and investor education, and positions its managed sandbox runtime as a turnkey solution for teams that want to deploy agents without building underlying infrastructure. Sandbox runtime is priced at 0.072 yuan per hour, one of the most transparent entry-level pricing packages in the domestic market.
2. Build vs Buy: Self-Hosted Harness vs Managed Intelligence
The central architectural decision for every enterprise adopting agent technology is whether to build and operate a self-hosted harness or consume a managed agent service. The two approaches differ across five core dimensions, and each carries distinct tradeoffs for control, operational cost and time-to-market.
2.1 Session History Management
In a self-hosted setup, engineering teams are responsible for building their own session storage layer, implementing message history persistence, and designing recall and rollback mechanisms. Every conversation state must be serialized, stored and reloaded manually when sessions resume. This gives full control over data format and retention policy, but adds significant engineering work.
Managed platforms handle persistence natively. Session history is durably stored on the platform side and can be retrieved at any time in a standardized format. This removes a major engineering burden, especially for teams supporting long-running, multi-session agent workflows.
2.2 Tool Execution
Self-hosted harnesses require teams to implement their own tool parsing, execution and result backfill logic. Every tool call emitted by the model must be intercepted, validated, executed in a controlled environment, and the output returned to the model in the correct schema. This allows full customization of tool behavior, but requires ongoing maintenance as new tools are added.
Managed services include built-in tool execution running on cloud-hosted sandboxes. For custom internal tools, platforms typically provide a standard registration interface for customers to connect their own endpoints, while the platform handles the invocation lifecycle, error handling and result formatting.
2.3 Sandbox Isolation
Running untrusted code and third-party tool outputs requires secure isolation. Self-hosted deployments require teams to provision, configure and maintain their own sandbox environments, including network policies, dependency management and resource limits. This is resource-intensive but gives full control over security boundaries.
Managed platforms provide pre-configured, isolated sandbox environments out of the box. Customers can configure network access rules and allowed dependency packages according to their security policies, without maintaining the underlying sandbox infrastructure.
2.4 Failure Recovery
Long-running agent workflows can fail for many reasons: network interruptions, tool timeouts, model errors or resource exhaustion. In self-hosted setups, teams must implement their own state checkpointing, retry logic and breakpoint resumption mechanisms.
Managed platforms handle failure recovery automatically. If execution is interrupted, the platform resumes from the last valid checkpoint without manual intervention. This is particularly valuable for long-duration tasks such as large document processing and multi-step data analysis.
2.5 Timeout and Retry Scheduling
Rate limiting, timeout policies and retry logic are critical for production reliability. Self-hosted teams must implement throttling, backoff and retry scheduling from scratch, tuned to their specific workload patterns.
Managed platforms operate centralized retry and scheduling systems that handle rate limits, transient errors and resource contention automatically. This reduces operational overhead and improves overall task success rates.
Practical Selection Criteria
Self-hosted harnesses are the preferred choice for two broad categories of use cases: internal enterprise business systems with strict data residency requirements, and workloads with extreme security and compliance constraints that demand full control over every layer of the stack.
Managed agent services excel for customer-facing products, teams that need fast time-to-market, and use cases where standard tooling and sandboxing are sufficient. They allow product teams to focus on agent design and business logic rather than infrastructure engineering.
For enterprise-grade deployments, four requirements are nearly universal: full controllability over execution, complete auditability of every action, the ability to pause and interrupt running tasks, and exportable session records. These apply regardless of whether the harness is self-hosted or managed.
For organizations that operate a mix of self-hosted and managed agents across multiple vendors, consistent governance becomes a central operational challenge. An enterprise-grade LLM gateway such as 4sapi can unify access across heterogeneous agent endpoints, centralize authentication and authorization, and provide consistent audit logging and usage metering. This gateway layer reduces integration complexity and ensures uniform security policies across every agent platform in use.
3. Kimi’s Capability Matrix: Pre-Built Skills, Plugins and Vertical Agent Templates
Kimi Hosted Agents launches with a broad pre-built capability set designed to reduce setup time for common enterprise use cases. The platform includes more than 140 built-in skills and over 30 native plugins, covering web research, document processing, code execution, office productivity and content generation. Customers can also build and register custom plugins to connect internal systems and proprietary data sources.
3.1 Data Access and Utility Plugins
The native plugin portfolio covers several high-demand categories:
- Web browsing and search: real-time web information retrieval, search engine integration and targeted site crawling.
- Document parsing: support for Word, PDF, Excel, PPT, images and other common formats, with structured extraction capabilities.
- Code interpreter: in-sandbox code execution for data analysis, chart generation and computational tasks.
- Productivity and collaboration tools: calendar integration, email operations and enterprise system connectors.
- Vertical industry tools: specialized plugins for financial data retrieval, regulatory document querying and industry-specific databases.
All plugins are available on both web and desktop clients, and can be invoked automatically by the agent or triggered explicitly by users.
3.2 Official Financial Agent Matrix
One of Kimi’s most distinct differentiators is its deep set of pre-built financial agent templates, covering the full investment and compliance workflow. The official agent matrix includes dedicated agents for securities research, report analysis, fund research, compliance review, investor education and fundamental analysis.
Each pre-built agent comes with tuned system prompts, configured tool sets and domain-specific prompt engineering, allowing financial institutions to deploy specialized agent workflows with minimal customization. This vertical focus addresses a key pain point for China’s financial services industry, where generic agents often lack domain accuracy and regulatory awareness.
3.3 Continuous Capability Expansion
Kimi states that its agent matrix and plugin library will be updated on an ongoing basis, with new industry templates and tool integrations added according to customer demand. The platform also supports fully custom agent development for use cases that do not fit standard templates.
4. Transparent Three-Part Pricing Model
Kimi Hosted Agents uses a three-component pricing structure designed to be predictable and itemized. The platform exposes both REST API and command-line CLI interfaces, and organizes its resource model around four core concepts:
- Agent: the static definition and configuration of an intelligent agent, including its instructions, tool set and behavioral parameters.
- Environment: the runtime execution environment, including installed dependencies, initialization scripts and versioned configuration.
- Session: a single task execution instance, with its own state, execution history and final output.
- Vault: centralized credential and key management, with secrets scoped to individual sessions to minimize exposure.
4.1 Sandbox Runtime Pricing
Sandbox usage is billed by runtime duration, with a standard configuration of 1 vCPU and 1 GB of memory priced at 0.072 yuan per hour. Billing is metered by actual running time, with a minimum chargeable unit per task. This pay-as-you-go model means customers only pay for the compute time they actually consume, with no upfront reservation fees.
4.2 Model Inference Pricing
Model token consumption generated during agent execution is billed separately according to standard model inference pricing. Detailed rate cards for each supported model are published on the platform’s pricing page, ensuring customers can estimate total cost in advance.
4.3 Plugin Usage Pricing
Plugin invocations are charged on a per-call basis, with tiered pricing for different plugin categories. Price points range from 0.001 yuan per call for general web search plugins, to higher rates for premium financial data plugins such as earnings report search, financial database queries and specialized market data feeds.
The tiered structure aligns cost with the value of the underlying data service, and allows teams to control spending by selecting only the plugins required for their use case.
4.4 Total Cost Structure
Total cost for any agent execution is the sum of three components: sandbox runtime fees, model inference fees and plugin invocation fees. With entry-level sandbox pricing starting at 0.072 yuan per hour and basic plugin calls starting at 0.001 yuan per invocation, the platform offers one of the lowest entry price points in the domestic managed agents market. Enterprise customers also have access to custom pricing plans for volume usage and dedicated resource allocation.
5. Differentiator: Kimi Inference Hook for Security Governance
Beyond standard runtime features, Kimi introduces a dedicated security control layer called the Kimi Inference Hook, embedded directly in the model inference pipeline. This component is designed to address one of the biggest concerns in enterprise agent deployment: ensuring that agent inputs and outputs comply with internal policies, regulatory requirements and security rules.
5.1 How the Inference Hook Works
The inference hook sits between the agent runtime and the underlying model, intercepting every input prompt and every model output before it reaches the next stage. Enterprise security teams can define custom policy rules that are enforced automatically at this layer, without modifying agent definitions or business logic.
Typical use cases include:
- Sensitive information redaction for input prompts and output responses
- Compliance policy validation against industry regulations
- Business rule enforcement to prevent out-of-scope agent actions
- Content moderation and harmful output filtering
Because the hook runs at the platform layer, policy updates take effect immediately across all agents in the organization, without requiring redeployment or changes to individual agent configurations.
5.2 Core Design Principles
The inference hook architecture is built around four design priorities:
- Enterprise-defined policies: Security teams retain full control over rule logic, rather than relying solely on platform default safeguards.
- Unified platform security capabilities: Common security functions such as content detection and data masking are implemented once at the platform level and applied consistently everywhere.
- Zero-downtime policy updates: Security rules can be updated and rolled out without interrupting running agent tasks.
- Platform-grade engineering reliability: The hook layer is built for high throughput and low latency, so security controls do not meaningfully degrade agent performance.
This built-in governance layer is a key advantage of managed agent platforms over fully self-hosted solutions. Building equivalent in-house security middleware would require significant engineering investment and ongoing maintenance.
6. Market Landscape: Multiple Players Converge on Agent Runtime
The entry of Kimi Hosted Agents intensifies competition in a market that is already evolving rapidly. Each major player brings different strengths and strategic positioning, reflecting the breadth of enterprise demand for agent technology.
OpenAI Agents API
OpenAI retains the broadest global ecosystem and the most complete feature set. Its native Codex integration, deep tool ecosystem and flexible deployment options make it the default choice for many international teams and software engineering use cases. The platform’s self-hosted sandbox option also appeals to organizations with strict data residency requirements.
Cursor Managed Agents
Cursor’s agent offering is deeply integrated with its code editor product and targets software development teams as its primary audience. Its open SDK and support for custom agent definitions give engineering teams high flexibility, and its developer-first positioning makes it popular among engineering organizations that want deep customizability.
Anthropic Claude Agent Ecosystem
Anthropic’s agent strategy is built on the reliability and long-context strength of its core models. The Claude Agent SDK emphasizes predictable, stable agent behavior for complex enterprise workflows, and is widely adopted for use cases that require high reasoning accuracy and low hallucination rates.
Kimi Hosted Agents
Kimi’s strongest positioning is in the China domestic market, with optimized Chinese language understanding, deep vertical financial templates and transparent, low-entry pricing. Its fully managed, turnkey model appeals to teams that want to deploy agents quickly without building infrastructure, particularly in the financial and professional services sectors.
Broader Industry Trend
Across all players, the industry is moving through a clear structural shift. Where previous generations of AI products competed on raw model benchmarks, the agent era competes on runtime infrastructure, orchestration capability, security governance and vertical ecosystem depth. The harness — the runtime environment that hosts, manages and secures running agents — has become the new core infrastructure layer.
For most large enterprises, the end state will not be a single agent platform. Instead, organizations will adopt a multi-platform strategy, selecting the best agent for each use case while maintaining centralized governance and security controls. This hybrid approach maximizes capability flexibility while managing compliance and cost.
Conclusion
The launch of Kimi Hosted Agents is a clear signal that the managed agents API market is moving from early exploration into mainstream production adoption. What began as a set of experimental developer tools is maturing into a standardized enterprise infrastructure category, with clear pricing models, security controls and vertical industry templates.
The core choice facing engineering teams — self-hosted harness versus managed service — will remain the central architectural decision for the foreseeable future. Self-hosted deployments offer maximum control and security, at the cost of higher engineering overhead. Managed services deliver faster time-to-market and lower operational burden, with built-in security and reliability features.
As the number of agent platforms grows, unified governance and interoperability will become increasingly important. Enterprises that build their agent architecture around standardized interfaces, centralized access control and defense-in-depth security will be best positioned to adopt new capabilities as they emerge, without being locked into a single vendor.
Over time, the agents API market will likely follow the same trajectory as earlier cloud service categories: initial fragmentation, followed by standardization of core interfaces, and eventually consolidation around a small set of dominant platforms with differentiated vertical capabilities. In the near term, however, the market will remain dynamic, with continued innovation in runtime architecture, security governance and industry-specific agent templates.
