甲骨文公司（Oracle）推出的全新 Fusion Claw 运行时，据分析师称，可能让客户对用于自动化业务流程的人工智能（AI）代理拥有更多控制权。但分析师也警告称，该技术仍处于早期采用阶段，客户可能会面临与甲骨文应用程序和云服务绑定更紧密的风险。早在今年三月，这家“大红公司”（Big Red）就曾承诺推出“能够推理、决策并为实现既定业务目标而行动的应用程序”，如今则推出了用于管理这些应用程序行为的运行时环境。
Oracle's new Fusion Claw runtime could give customers more control over AI agents automating business processes, analysts say. But they warn that adoption remains at an early stage and customers risk becoming more tightly tied to Oracle's applications and cloud. Having promised "applications that can reason, decide, and act in pursuit of defined business objectives" back in March, Big Red has now introduced a runtime to govern their actions, running on its cloud and integrated with its Fusion applications.
该运行时部署在甲骨文的云平台上，并与 Fusion 应用程序集成。甲骨文将 Fusion Claw 的目标市场锁定在 ERP、财务和供应链管理领域，适用于运行于甲骨文云基础设施（OCI）上的云端 Fusion 应用程序。该技术将由前沿模型提供支持，包括来自谷歌和 OpenAI 的模型，未来还将逐步支持其他模型。甲骨文首席执行官 Mike Sicilia 表示，公司正努力解决如何从人工智能投资中“为企业创造可衡量的价值”这一问题。
Oracle is targeting ERP, finance, and supply chain management with Fusion Claw for cloud-based Fusion applications running on Oracle Cloud Infrastructure (OCI). It will be powered by frontier models including those from Google and OpenAI, with support for others planned over time. Oracle CEO Mike Sicilia said the company was trying to address the problem of getting "measurable value for our business" from AI investment.
他表示：“Fusion Claw 通过将人工智能从辅助工具升级为执行工具，帮助回答了这个问题，使客户能够在最重要的业务流程中自动化日益复杂的工作，同时具备企业级治理能力。”分析师们对此举措表示欢迎，认为这为客户在甲骨文 Fusion 云应用程序中部署 AI 代理提供了控制权和保障。Gartner 最近对 Fusion Claw 的“初步评估”指出，该技术提供了一种方法，可以在“明确的策略限制内运行隔离容器，以执行复杂的供应链优化和子总账对账任务”，而无需依赖本地桌面脚本或聊天机器人提示。
"Fusion Claw helps answer that question by moving from AI assistance to execution, enabling customers to automate increasingly complex work across their most important business processes, with enterprise-grade governance," he said. Analysts welcomed the move as a path that offers customers control and assurance in their AI agent implementations on Oracle Fusion Cloud Applications. A recent Gartner "first take" on Fusion Claw said the technology was a way to provision "isolated containers to execute heavy supply chain optimizations and subledger reconciliations within clear policy limits" rather than running local desktop scripts or chatbot prompts.
Gartner 表示：“甲骨文没有像以往那样将人工智能作为附加的副驾驶或用户界面封装层，附加到传统的交易型后端系统上，而是正在围绕原生代理式工作流程重构其应用程序基础架构。”
"Instead of bolting AI onto legacy transactional back ends as a sidecar copilot or UI wrapper, Oracle is rebuilding its application foundation around native agentic workflows," Gartner said.
这位分析师表示，其他供应商应当重视这一举措，因为客户需要摆脱基于提示的界面和仅能处理常规数据录入及基础任务自动化的“侧边聊天机器人”，这些工具带来的“边际收益递减”，无法证明企业投入AI的合理性。Gartner指出：“AI算力应聚焦于那些此前在核心ERP系统中无法解决的最复杂的运营挑战，例如多级供应链可见性、网络设计、销售与运营规划（S&OP）、物流路线规划以及负载优化。”
Other vendors should take heed of the move, the analyst said, because customers need to get away from prompt-based UIs and sidecar chatbots that only address routine data entry and basic task automation and offer "diminishing returns" that fail to justify enterprise AI spending. "AI horsepower should be pointed at the most complex operational challenges that were previously unsolvable inside core ERP systems, such as multi-echelon supply chain visibility, network design, sales and operations planning (S&OP), logistics routing, and load optimization," Gartner said.
然而，用户若想部署这类企业级应用，则会面临成本障碍，因为使用大型语言模型（LLM）处理成千上万甚至数百万条“交易记录、明细项或约束变量……在企业规模下很快就会变得在经济上不可行。”Gartner表示。
However, users trying to get these kinds of enterprise-scale applications hit a cost barrier because using an LLM to address thousands or millions of "transactional records, line items, or constraint variables… quickly becomes financially unviable at enterprise scale," it said.
Oracle通过将推理与计算执行分离来规避这一问题。这位分析师称，一个智能体只需评估上下文、整合工具并定义执行逻辑一次，而计算和多变量优化则被委托给确定性计算引擎，“从而避免将数百万条记录传入昂贵的LLM上下文窗口”。隔离的容器还提供了特定的权限限制、业务对象状态和政策指南。Gartner表示，执行结束后，系统会留下提示，以跳过冗余路径并减少令牌消耗。
Oracle gets around the problem by separating reasoning from compute execution. An agent evaluates context, assembles tools, and defines execution logic just once, the analyst said. Calculations and multivariable optimizations are delegated to deterministic compute engines, "avoiding passing millions of records through expensive LLM context windows." Isolated containers also offer specific authority limits, business object states, and policy guidelines. After execution, the system leaves hints to skip redundant paths and cut token consumption, Gartner said.
Oracle已推出多款现成应用，包括用于提升财务总账准备度、制定人员配置计划以及整合运输流程的应用。Gartner建议，在人工智能智能体部署相关的招标请求（RFP）中纳入Fusion Claw，不过企业应用研究主管Patrick Connaughton警告称，Fusion Claw“目前仍处于非常早期的试点阶段”。另一家分析公司Forrester也欢迎Oracle推出Fusion Claw，该产品提供按用户或员工计费的定价模式，以及基于使用量的计费模式。
Oracle is offering ready-made applications, including ones to improve readiness of financial ledgers, make staffing plans and consolidate shipping. Gartner recommends including Fusion Claw in requests for proposals (RFPs) for AI agent deployments, although Patrick Connaughton, chief of research, enterprise applications, warned that Fusion Claw was "very much in early piloting mode still." Fellow analyst company Forrester also welcomed Oracle's introduction of Fusion Claw, which is offered under per-user or per-employee pricing as well as consumption-based models.
首席分析师克雷格·勒克莱尔（Craig Le Clair）表示：“商业应用与科技行业的激进布局保持一致。在我们看来，在智能代理系统能够被信任独立于设计者意图地创建工具和流程模式之前，我们还有许多信任问题需要解决。”然而，与Fusion Claw仅依赖Oracle的世界观不同，企业拥有混合环境，其中AI代理被构建在Salesforce和Oracle等云域中，流程平台如Pega或UiPath，以及框架如LangChain也广泛使用。
Principal analyst Craig Le Clair said: "Commercials are in line with the aggressive positioning from the tech industry. In our view, we have many trust issues to solve before agentic systems can be trusted to create tools and process patterns independently of designer's intent." However, in contrast with Fusion Claw's Oracle-only view of the world, enterprises have hybrid environments with AI agents built within cloud domains such as Salesforce and Oracle, process platforms like Pega or UiPath and frameworks like LangChain.
他表示：“像Oracle这样的云域最适合处理那些通用且不具备差异化功能的任务，这些是云域本身擅长的领域；而更复杂、更具变革性的定制代理则应着眼于专门的智能代理框架和超大规模云服务商平台。”锁定效应是一个问题。尽管Oracle的版本受到OpenClaw的启发，但它却是专有技术，并与Fusion平台紧密耦合，即运行在Oracle云基础设施上的Oracle Fusion Cloud应用程序。
"Cloud domains like Oracle are best suited for more generic, non-differentiating functions that the cloud domain knows well, while more complex and transformative custom agents should look toward custom agentic frameworks and the hyperscaler platforms," he said. Lock-in is an issue. While inspired by OpenClaw, Oracle's version is proprietary and tightly coupled to the Fusion platform, Oracle Fusion Cloud Applications running on Oracle Cloud Infrastructure.
“企业架构师是避免‘锁定效应’的关键，他们需要根据具体情况进行指导，例如在何时选择域云而非更开放的顶层平台，并考虑数据驻留和外部集成需求等因素。如果该云域的优势足够显著，锁定效应或许可以接受。”他表示。那些已在Oracle上投入大量资源的企业，可能会欢迎更严格的治理结构以及潜在的更低运行成本。
"Enterprise architects are the key to avoid 'lock-in' with guidance on when to select a domain cloud over a more open platform above, pointing to data residency and external integration needs, for example. If the domain strength is strong enough, lock-in may be OK," he said. Organizations already heavily invested in Oracle may welcome tighter governance and potentially lower running costs.
而使用更多样供应商组合的企业，则可能认为与Fusion和OCI的绑定较难获得合理 justification。®
Those using a broader mix of suppliers may find the ties to Fusion and OCI harder to justify. ®