Managed Agent Services Go Commercial: Why the Agent Runtime Is Becoming a Separate Billing Unit

    This article examines the commercialization of managed agent services and why the agent runtime is becoming a standalone billing object. It explains the role of a harness for long-running tasks, breaks down a three-part pricing model covering sandbox runtime, model inference, and plugin calls, and offers practical guidance on cost estimation, sandbox lifecycle management, caching, model routing, and budget monitoring. It also covers enterprise deployment patterns, security controls, and the broader industry trend toward decoupling runtime, tools, and inference.

    1. Background: From Model APIs to Managed Agents

    Over the past two years, the main way developers have accessed large models has been through inference APIs billed by token usage. This model suits single-turn or short-chain conversational tasks, but problems emerge when businesses move into the agent stage. Agent tasks often run for tens of minutes or even days and involve multi-step tool calls, state preservation, and exception recovery. An inference API alone cannot cover the full execution chain.

    As a result, leading model vendors have begun offering the runtime and orchestration system built around the model—commonly referred to in the industry as a Harness—as an independent cloud-managed service. Recently, a vendor officially commercialized this type of managed agent service, providing enterprises and developers with a 24/7 managed runtime environment. This move signals that the agent runtime is becoming an independent billing object, a development worth attention from every team building agents.

    2. What Is a Harness? The Technical Foundation for Long-Running Tasks

    A Harness refers to the runtime and orchestration system built around a large model. It sits above the model inference API and typically handles the following responsibilities:

    1. Model call orchestration: organizing multi-round inference and controlling context windows and prompt assembly.
    2. Tool and plugin execution: scheduling external tool calls and managing parameter validation and result injection.
    3. Session and state management: maintaining session history, task context, and intermediate artifacts.
    4. Task scheduling and recovery: supporting failure recovery, timeout retries, and resumption from checkpoints. This is key to whether long-running tasks can execute reliably.
    5. Permission and security controls: integrating permission checks, sensitive data blocking, and related capabilities.

    For tasks that need to run continuously across hours or days—such as investment research monitoring or operations response—this layer is almost essential. Without it, developers would need to build message queues, persistent storage, retry mechanisms, and sandbox environments themselves, resulting in considerable engineering cost.

    3. Billing Model: How the Three Cost Components Fit Together

    The pricing formula disclosed in this commercialization is representative. Total cost is the sum of three parts:

    1. Total cost = sandbox runtime duration × unit price
    2. + model inference cost (billed by token)
    3. + Σ(plugin call count × plugin unit price)

    The components are as follows:

    1. Cloud sandbox fee: Agents need an isolated runtime environment to execute code and access files, billed by runtime duration. Based on the official example, a sandbox billed continuously for 30 days would cost about CNY 51.84 in duration fees (CNY 0.072 per hour), excluding inference and plugin costs. In real workloads, sandboxes are often started and stopped on demand, so estimates should account for task activity.

    2. Model inference fee: This follows the traditional token-based billing logic. Multi-round agent inference amplifies token consumption, especially for long-context tasks.

    3. Plugin call fee: Plugin unit prices range from CNY 0.001 to CNY 0.45 per call. For example, web search costs approximately CNY 0.03 per call, image search approximately CNY 0.045 per call, and some financial data plugins cost CNY 0.35 or CNY 0.45 per call.

    This pricing structure sends an important signal: the cost model for agent services is no longer a single “token cost,” but a composite model of “compute environment cost + inference cost + data/tool cost.”

    4. Implications for Developers: How to Estimate and Optimize Costs

    Based on this billing structure, teams should consider the following practices in architecture design:

    1. Start and stop sandboxes on demand: Long-running tasks do not need a sandbox to stay online continuously. Tasks can be split into a state machine of trigger → execute → persist → sleep, with persistent storage holding intermediate state. This can significantly reduce duration fees.
    2. Cache plugin calls: For high-frequency plugins whose results change slowly, such as search and market snapshots, implement result caching and deduplication at the application layer to avoid duplicate charges.
    3. Tier inference costs: Use small models for simple judgments and large models for complex planning. Model routing can reduce overall inference overhead.
    4. Budget and monitoring: Report costs for each agent instance and combine log services with monitoring and alerting to prevent runaway loops.

    An illustrative cost estimation pseudocode example:

    1. def estimate_cost(sandbox_hours, tokens, plugin_calls):
    2. cost = sandbox_hours * 0.072
    3. cost += tokens * price_per_token
    4. for plugin, n in plugin_calls.items():
    5. cost += n * plugin_unit_price[plugin]
    6. return cost

    5. Enterprise Deployment Scenarios and Security Controls

    Based on disclosed information, typical enterprise applications of managed agents fall into two categories. One integrates with internal systems for tasks such as business monitoring and operations response. The other is embedded in external products to provide services such as investment research assistants and customer service assistants. The first templates focus on financial scenarios, covering investment research modeling, IPO tracking, position monitoring, regulatory case daily reports, fund reconciliation, and similar tasks. The offering provides more than 100 skills and dozens of plugins, with data sources covering mainstream financial data services.

    Notably, these services are increasingly making security components independent: before each model inference, they inspect task inputs and content returned by tools, and allow or block sensitive data according to enterprise-defined rules. When integrating, enterprises should focus on three areas. First, the data boundary: clarify which fields are allowed to enter the model context. Second, least-privilege tool invocation: an agent should hold only the minimum permissions needed to complete the task. Third, audit log completeness: all inference and plugin calls should be traceable.

    6. Industry Trend Observations

    Some international vendors have already commercialized agent services with per-conversation pricing. The vendors in China discussed here have adopted a composite billing model of “duration + tokens + number of calls,” reflecting a shared judgment: the agent infrastructure layer has independent commercial value. For developers, this means two things:

    1. When selecting a solution, build a multidimensional cost model. Do not look only at model unit price. Evaluate total cost of ownership across sandbox, plugins, and inference.
    2. A new trade-off has emerged between self-hosting and managed services. For small task volumes, self-hosting a Harness plus inference APIs may be more cost-effective. For tasks that require highly available long-running execution, managed services can significantly reduce engineering burden.

    The infrastructure for the agent era is taking shape quickly. Changes in billing are only the surface; the underlying reality is that runtime environments, tool ecosystems, and model inference are decoupling into three independent technical and commercial layers. Understanding this is essential for sound technology selection and cost planning.